GLM Coding Plan vs Claude Code: What Each Actually Costs (2026)

By Chris Chen · Updated September 24, 2026

Z.ai’s GLM Coding Plan starts at $18 a month and runs inside Claude Code. Anthropic’s own Max 5x plan is $100 a month. That 5.5x gap is the reason this comparison keeps coming up — and it is also the reason almost every write-up of it is wrong, because the two products do not bill for the same thing, do not count tokens the same way, and do not restrict you in the same places.

This page does the boring work: every number below is read verbatim from a vendor’s own pricing or documentation page on 30 August 2026, and every calculation built on top of those numbers is labelled as our arithmetic rather than presented as a vendor figure. Where a vendor does not publish something, we say so instead of filling the gap with a third-party estimate.

The short answer

• If your bill is the constraint, the GLM Coding Plan is dramatically cheaper per unit of work, and the gap is wider than the headline prices suggest for developers outside Asia — see the peak-hours section below.
• If cache-heavy agent loops are your workload, one number cuts the other way: Claude Sonnet 5 reads cached input at $0.20 per million tokens, while GLM-5.3 charges $0.26. On a workload that is 95% cache hits, Anthropic is the cheaper cache.
• If you need to keep the tool and swap the engine, that is exactly what the GLM Coding Plan is built for — Z.ai publishes a Claude Code endpoint for it.
• If uninterrupted throughput matters more than price, the flat plans on both sides run out, and both hide the exact cut-off. Neither vendor publishes a hard token number you can plan against.

What each side actually charges

Anthropic subscriptions (support.claude.com, read 30 August 2026)

Plan	Price	Billing
Free	$0	—
Pro	$20/month or $200/year	Monthly or annual
Max 5x	$100/month	Monthly only
Max 20x	$200/month	Monthly only

Anthropic’s help centre states that Max prices are for web subscriptions and that mobile pricing may vary by app platform. It also states that Max plans carry both a five-hour session limit and a weekly limit that applies across all models, with a fixed weekly reset time assigned per account. Anthropic does not publish a token figure for either window. We have covered what it does publish in detail on Claude Pro usage limits and Claude Max 5x vs 20x.

Anthropic API rates, per million tokens (platform.claude.com, read 30 August 2026)

Model	Input	5-min cache write	1-hour cache write	Cache hit	Output
Claude Opus 5	$5	$6.25	$10	$0.50	$25
Claude Sonnet 5	$2	$2.50	$4	$0.20	$10
Claude Sonnet 4.6	$3	$3.75	$6	$0.30	$15
Claude Haiku 4.5	$1	$1.25	$2	$0.10	$5

Anthropic’s Batch API takes 50% off both input and output, which puts Sonnet 5 at $1 / $5. Claude 4.6 and later models include the full one-million-token context window at standard per-token pricing, so a 900k-token request is billed at the same rate as a 9k one.

Z.ai API rates, per million tokens (docs.z.ai, read 30 August 2026)

Model	Input	Cached input	Cached input storage	Output
GLM-5.3	$1.40	$0.26	Limited-time free	$4.40
GLM-5.2	$1.40	$0.26	Limited-time free	$4.40
GLM-5.3-Flash	$0.15	$0.03	Limited-time free	$0.50

Updated 13 September 2026: the 50% launch discount on GLM-5.3-Flash has been withdrawn and the table above now carries Z.ai’s list prices, re-read on the official pricing page that day. The discount was still live on 10 September, three days after the expiry Z.ai had published for it; see the 13 September update at the foot of this page for the dates and for what the reversion does and does not change. GLM-5.3 itself is priced identically to GLM-5.2 — the new generation did not come with a posted price rise.

The GLM Coding Plan (docs.z.ai, read 30 August 2026)

This is the product people actually mean when they compare GLM to Claude Code. Z.ai’s overview page says it starts “at just 18 USD per month” and lists three individual tiers — Lite, Pro and Max. The per-tier dollar prices for Pro and Max are not published on the documentation pages we read; they appear only on the subscription page inside a Z.ai account. We are not going to reprint a third-party number for them. What Z.ai does publish, and what actually determines your ceiling, is the credit allowance:

Tier	5-hour credits	Weekly credits
Lite	2,000	10,000
Pro	12,000	60,000
Max	28,000	140,000

Five-hour credits refresh dynamically — the quota resets five hours after it is consumed, rather than on a fixed clock. Weekly credits activate when you subscribe and reset every seven days from that moment. Both plans support GLM-5.3 and GLM-5.3-Flash only; Z.ai states that requests for GLM-5.2 or GLM-5.1 are automatically routed to GLM-5.3, and GLM-4.7 requests to GLM-5.3-Flash.

Credits are not tokens. Z.ai publishes the conversion explicitly:

Credit usage = (input tokens x input multiplier + cached input tokens x cached multiplier + output tokens x output multiplier) / 10,000

Product	Input multiplier	Cached input multiplier	Output multiplier
GLM-5.3	6.9	1.7	24
GLM-5.3-Flash (incl. vision MCP)	2.3	0.56	8
Web Search / Web Reader / Zread MCP	—	—	1.2 per call

Three reasons the per-token prices are not comparable

1. Claude Sonnet 5 uses a different tokenizer than Claude Sonnet 4.6

This is buried in Anthropic’s pricing documentation and it is the single most under-reported cost fact of 2026. Verbatim: Claude 4.7 and later models use a newer tokenizer that “produces approximately 30% more tokens for the same text,” and Claude Sonnet 4.6 and earlier use the previous tokenizer. Anthropic notes the exact increase depends on content and workload shape.

Our arithmetic on that official statement: if the same source file becomes roughly 30% more tokens on Sonnet 5 than it would have on Sonnet 4.6, then Sonnet 5’s headline $2 per million input tokens behaves more like $2.60 per million old-tokenizer-equivalent tokens when you are comparing against a model on a different tokenizer. The move from Sonnet 4.6 at $3 to Sonnet 5 at $2 is still a real price cut on the same text — roughly 13% by our arithmetic, not the 33% the sticker suggests.

Z.ai does not publish a tokens-per-character figure for GLM-5.3, and we could not find an official statement about how its tokenizer compares to either Anthropic generation. So the honest conclusion is not “GLM is 30% cheaper than the table says” — it is that cross-vendor per-token prices cannot be compared at all. Compare bills for a job, not prices for a token. That is the same unit problem we set out on AI coding assistant pricing.

2. The two vendors charge for caching in opposite shapes

Anthropic charges a premium to write to cache (1.25x base input for a five-minute cache, 2x for a one-hour cache) and then 0.1x base input to read it. Z.ai charges a flat cached-input rate and lists cached-input storage as free for a limited time — there is no separate write premium in its published table.

The consequence surprises people: on Sonnet 5 a cache hit costs $0.20 per million tokens; on GLM-5.3 it costs $0.26. For a coding agent that re-reads the same large repository context dozens of times per session — which is the normal shape of Claude Code work, as we set out in Claude Code’s context window — the dominant line item on the bill is cache reads, and on that line Anthropic is 23% cheaper by our arithmetic. GLM’s advantage lives in output tokens and in uncached input, not in cache.

3. One side sells tokens, the other sells credits with a time-of-day discount

Z.ai’s documentation states that during off-peak hours model usage is charged at 50% of the standard credit rate, and defines peak hours as Monday to Friday, 14:00–18:00 Singapore Standard Time (UTC+8). There is no equivalent on the Anthropic side. This turns out to matter more than any headline price, and it gets its own section below.

Same job, two bills: a worked example

Nobody buys a token. People buy a week of work. So here is one week of agentic coding, defined explicitly so you can substitute your own numbers. The workload below is our assumption, not a vendor figure — it is the shape agentic coding tends to take: an enormous amount of repeatedly re-read context, and a comparatively tiny amount of generated output.

• 40 million input tokens read over the week
• Of those, 38 million are cache hits and 2 million are fresh writes (a 95% cache hit rate)
• 1.2 million output tokens

Claude Sonnet 5 on the API

Line	Our arithmetic	Cost
Cache hits	38M x $0.20/M	$7.60
Cache writes (5-min)	2M x $2.50/M	$5.00
Output	1.2M x $10/M	$12.00
Week		$24.60
Month (4.33 weeks)		~$107

That is the first genuinely useful result: at this workload, paying Anthropic per token costs roughly what Max 5x ($100/month) costs as a flat plan. The flat plan is not a discount at this volume — it is insurance against a spike, plus access to Claude Code without wiring up API keys.

GLM-5.3 on the API

Line	Our arithmetic	Cost
Cache hits	38M x $0.26/M	$9.88
Uncached input	2M x $1.40/M	$2.80
Output	1.2M x $4.40/M	$5.28
Week		$17.96
Month (4.33 weeks)		~$78

Roughly 27% cheaper than the Anthropic API on the same token counts, by our arithmetic — a real gap, but nothing like the 5.5x the $18-versus-$100 headline implies. Notice where the gap comes from: the output line ($5.28 versus $12.00) and the uncached input line. The cache line is worse on GLM.

The same job on the GLM Coding Plan

Now run the identical workload through Z.ai’s published credit formula. All of the following is our arithmetic on Z.ai’s official multipliers.

Line	Calculation	Weighted tokens
Cached input	38M x 1.7	64.6M
Uncached input	2M x 6.9	13.8M
Output	1.2M x 24	28.8M
Total	sum / 10,000	10,720 credits

At the standard (peak) credit rate that week costs 10,720 credits — which is just over the Lite tier’s 10,000 weekly cap and comfortably inside Pro’s 60,000. At the off-peak rate it halves to 5,360 credits, which fits inside Lite with 46% of the weekly allowance still unused.

So the same week of work costs approximately $107 on the Anthropic API, $78 on the Z.ai API, or one $18 Lite subscription if your hours fall off-peak. That is the real comparison, and it is a wide gap — just not the gap the headline numbers describe, and it depends entirely on when you work.

The peak-hours discount most Western developers get by accident

Z.ai defines peak hours as Monday to Friday, 14:00–18:00 Singapore Standard Time. Singapore is UTC+8. Our arithmetic on that official definition:

Zone	Z.ai peak window falls at	Practical effect
UTC	06:00–10:00	Overlaps the European morning
US Eastern (UTC-4 in August)	02:00–06:00	The middle of the night
US Pacific (UTC-7 in August)	23:00–03:00	The middle of the night
Central Europe (UTC+2 in August)	08:00–12:00	Directly on the working morning

A developer in North America who works normal hours is, by our arithmetic, off-peak essentially 100% of the time — which means the effective credit rate is half the posted rate and the token allowance is double. A developer in Berlin or Paris hits the peak window every single working morning and gets the full rate on a meaningful share of their week. A developer in Singapore, Hong Kong or Shenzhen is squarely in it.

This is the sort of detail that quietly decides whether a plan is cheap or not, and we have not seen it stated anywhere else. It also explains the range in Z.ai’s own published allowance table, which we reproduce here because it is the only official token-level guidance either vendor gives:

Cache hit rate	Model	Lite (M tokens/week)	Pro	Max
95%	GLM-5.3	48–97	290–580	676–1,352
95%	GLM-5.3-Flash	146–292	877–1,755	2,047–4,095
98%	GLM-5.3	52–104	313–627	731–1,463
98%	GLM-5.3-Flash	158–317	950–1,900	2,217–4,433

Z.ai states the range is constructed exactly the way we described: the top of each range assumes all usage is off-peak at 0.5x credits, the bottom assumes all usage is at peak at 1x. In other words, the entire spread between the low and high figure in that table is the time zone you happen to live in.

Z.ai also claims that fully exploiting off-peak pricing saves “up to 92%” versus pay-as-you-go GLM-5.3 API calls. That is a vendor claim, not something we have reproduced, and “up to” is doing work in that sentence.

Where the GLM Coding Plan is more restrictive than Claude Code

Price is not the only axis, and Z.ai’s own documentation is unusually blunt about the constraints. All of the following is stated on docs.z.ai:

• Approved tools only. The plan is “strictly limited to use within officially supported tools and products.” Claude Code, Cline and OpenCode are named. Using the quota outside those tools is prohibited by the terms, not merely unsupported.
• Two models only. Plan traffic can call GLM-5.3 and GLM-5.3-Flash. Older model IDs are silently re-routed rather than served.
• You must point at a specific endpoint. For Claude Code the base URL is https://api.z.ai/api/anthropic; other tools use https://api.z.ai/api/coding/paas/v4. Z.ai’s own FAQ says the most common support ticket — an “insufficient balance” error despite an active subscription — is caused by not doing this.
• No overflow onto your balance. When plan quota runs out, calls simply stop until the next five-hour refresh; Z.ai will not silently bill your account balance. That is a genuine consumer protection and the opposite of how API overage usually works — but it also means a deadline can hit a wall you cannot pay your way past.
• Non-refundable. Z.ai states subscriptions cannot be refunded once purchased, used or not.
• Same-tier upgrades do not take effect immediately. Moving from a Lite monthly to a Lite annual plan stacks validity periods (Z.ai’s own example: 13 months total) rather than switching you over. Cross-tier upgrades do take effect immediately, with the remaining value pro-rated into account balance.

Anthropic’s constraints are looser in kind but vaguer in specification: no approved-tool list, no endpoint gymnastics, but also no published number at all for what a five-hour or weekly window actually contains. We keep a running record of everything Anthropic does publish on when Claude’s usage limits reset.

The Claude side just got cheaper, permanently

One thing changed in Anthropic’s favour while we were checking these pages, and it is worth stating plainly because a lot of budget spreadsheets were built on the old assumption. Anthropic’s pricing documentation now says, verbatim, that the $2/$10 per million input/output token pricing for Claude Sonnet 5 — announced at launch as introductory pricing through 31 August 2026 — is now the standard price, and the previously scheduled increase to $3/$15 on 1 September 2026 will not occur.

If you had modelled a 50% input and output price rise landing this week, delete it. Our arithmetic on the worked example above: at $3/$15 that week would have cost $7.60 in cache hits plus $7.50 in cache writes plus $18.00 in output, or $33.10 — roughly 35% more than the $24.60 it actually costs. The cancellation is worth about $37 a month at that workload, which is a material change to the GLM comparison rather than a footnote.

Which one should you actually buy

Situation	What the numbers say
Occasional coding, a few hours a week	Claude Pro at $20/month. The GLM Lite plan is $18 and you would be buying a second tool to save $2.
Heavy agentic coding, North America or Latin America, budget-constrained	GLM Coding Plan. Your working hours are entirely off-peak, so the effective allowance doubles. This is where the gap is genuinely large.
Heavy agentic coding, Europe or Asia	Still cheaper on GLM, but run the credit arithmetic above against your real cache hit rate before assuming Lite is enough — peak-hour work burns credits twice as fast.
Cache-dominated workload above ~97% hit rate	The gap narrows sharply, because GLM’s cached input is more expensive than Sonnet 5’s. Compare bills, not stickers.
You cannot tolerate hitting a wall mid-deadline	Anthropic Max 20x, or the Anthropic API with no cap at all. GLM’s plan will not spill over onto your balance by design.
Compliance or data-residency constraints	Read both vendors’ terms before price enters the conversation. This page is a cost analysis and nothing more.

The option most people overlook is not either/or. Both products are configured per-project, so running Claude Code against Anthropic for architecture and review work and against Z.ai’s endpoint for bulk refactoring is a supported configuration on both sides. We walk through the full set of swap-the-engine options on Claude Code alternatives, and the head-to-head against OpenAI’s tool on Codex vs Claude Code.

What we could not verify

• GLM Coding Plan Pro and Max monthly prices. Z.ai’s documentation says only “starting at 18 USD per month.” Widely circulated third-party figures disagree with each other, so we have printed none of them. Check the subscription page in your own Z.ai account.
• What a Claude five-hour or weekly window contains in tokens. Anthropic has never published it, and the guidance in its Claude Code documentation is still expressed in hours of older model generations.
• GLM-5.3’s tokenizer ratio. Not published, so no honest cross-vendor per-token comparison is possible.
• Independent coding benchmarks for GLM-5.3 versus Claude Sonnet 5. Both vendors make performance claims; we do not restate vendor benchmark numbers as findings, and we have not seen a third-party reproduction we would stand behind.

Frequently asked questions

Can the GLM Coding Plan really run inside Claude Code?

Yes. Z.ai publishes a dedicated Anthropic-compatible endpoint (https://api.z.ai/api/anthropic) for Claude Code and Goose, and names Claude Code first in its list of supported tools. You are using Anthropic’s client with Z.ai’s models behind it.

Is GLM-5.3 cheaper than Claude Sonnet 5?

On output tokens and uncached input, yes, substantially — $4.40 versus $10 for output. On cached input, no: $0.26 versus $0.20. Which one wins your bill depends on your cache hit rate, and for agentic coding that number is usually very high.

Did Claude Sonnet 5 go up in price on 1 September 2026?

No. Anthropic’s pricing page states the scheduled increase to $3/$15 will not occur and that $2/$10 is now the standard rate.

What happens when GLM Coding Plan credits run out?

Calls stop until the next five-hour refresh. Z.ai states explicitly that it will not deduct from your account balance, and that plan subscribers cannot make API calls outside the plan quota within supported tools.

Is the GLM-5.3-Flash discount permanent?

No. Z.ai’s pricing page states the 50% promotion ends at 24:00 on 9 September 2026, Singapore time, after which Flash reverts to $0.15 input and $0.50 output per million tokens.

Read next

• AI coding assistant pricing — every published subscription tier, side by side
• Claude Code alternatives — including the keep-the-tool, swap-the-model category
• Codex vs Claude Code — the other comparison people run before buying
• Claude Max 5x vs 20x — if you have decided to stay on Anthropic

All vendor figures on this page were read from Anthropic’s and Z.ai’s own pricing and documentation pages on 30 August 2026. Prices change; check the source before committing a budget.

Update, 4 September 2026: the plan most write-ups describe is no longer on sale

Re-reading Z.ai’s own documentation today turned up something that changes how you should read almost every GLM Coding Plan comparison on the web, including the numbers people quote at us in email. On 30 July 2026, Z.ai published a Plan Update Announcement stating that the GLM Coding Plan moved to a credits-based system and that the previous plans are no longer sold to new users. Existing subscriptions were grandfathered: Z.ai calls them Legacy Plan V1, Legacy Plan V2 and Team Plan, and says their “price, benefits, usage limits, and calculation method stay unchanged until the end of the current billing cycle.”

That matters because of where the familiar numbers now live. The widely-quoted allowance figures — roughly 80 prompts per five hours on Lite, 400 per week, 400 and 2,000 on Pro, 1,600 and 8,000 on Max — sit on that announcement page under a heading Z.ai labels “Usage Reference for Legacy Plans.” They describe a plan a new subscriber cannot buy. If you signed up after 30 July 2026, your allowance is denominated in credits, not prompts, and no conversion between the two is published. We have not found a single third-party comparison that makes this distinction.

What one credit is actually worth — and why it is three different numbers

Z.ai publishes the credit formula in full: model credit usage equals (input tokens × input multiplier + cached input tokens × cached input multiplier + output tokens × output multiplier) ÷ 10,000. For GLM-5.3 the official multipliers are 6.9 input, 1.7 cached input, 24 output. Divide through and you get the credit price of a million tokens (our arithmetic):

GLM-5.3, per 1M tokens	Credits (our arithmetic)	Z.ai API list price	Implied $ per credit (our arithmetic)
Input	690	$1.40	$0.00203
Cached input	170	$0.26	$0.00153
Output	2,400	$4.40	$0.00183

The right-hand column is the point. If credits were simply a rebranded currency, that column would be one number repeated three times. It is not. A credit spent on cached input redeems for about $0.00153 of list-price service; a credit spent on plain input redeems for about $0.00203, a third more. In other words the credit table is not a linear rescale of the API price list — it quietly re-prices the model, and it re-prices it against exactly the token category that dominates agent work.

The cache-discount ladder

Earlier in this article we compared the absolute cache-hit prices — $0.20 per million on Claude Sonnet 5 against $0.26 on GLM-5.3. The sharper comparison is the ratio: how much cheaper is a cached token than a fresh one, on each of the three price lists a Claude Code user might actually be billed under?

Price list	Cached input vs fresh input	Source of the figure
Anthropic API (all models except Fable 5.1 / Mythos 5.1)	10x cheaper	Official: cache read is stated as 0.1x base input price
Z.ai API, GLM-5.3	5.4x cheaper	Our arithmetic: $1.40 ÷ $0.26
GLM Coding Plan credits, GLM-5.3	4.1x cheaper	Our arithmetic: 6.9 ÷ 1.7

Read down that column. Anthropic gives you a tenfold discount on cache reads. Z.ai’s public API gives you about 5.4x. And the subscription plan — the $18-a-month product that is the whole reason people compare GLM to Claude Code in the first place — gives you about 4.1x, the weakest of the three. The plan is stingier about caching than Z.ai’s own pay-as-you-go API is.

That is not a rounding artefact, and it points the same direction as everything else on this page. A Claude Code session re-reads the same repository context over and over; Z.ai’s own allowance table assumes cache hit rates of 95% to 98% precisely because that is the shape of the work. On a workload that is 95% cache reads, the category where GLM’s plan is least generous relative to its own API is the category that is nearly the whole bill. We wrote about why that context keeps getting re-read in how Claude Code’s context window actually behaves.

The same asymmetry shows up on output. On the API, GLM-5.3 output costs 3.14x its input ($4.40 ÷ $1.40). In credits, output costs 3.48x input (24 ÷ 6.9) — our arithmetic in both cases. Output is relatively dearer inside the plan too. What the plan is generous about, relative to the API, is plain uncached input: the one category a coding agent generates least of.

Checking our own arithmetic against Z.ai’s allowance table

A derivation like that is worth testing against something the vendor published independently. Lite carries 10,000 weekly credits. Take a workload that is 95% cache hits on input, and hold output aside for a moment: a million tokens then costs (0.05 × 6.9) + (0.95 × 1.7) = 1.96, times 100, or about 196 credits per million (our arithmetic). Ten thousand credits therefore buys roughly 51M tokens a week at the peak rate and about 102M at the off-peak half rate.

Z.ai’s own published figure for Lite at a 95% cache hit rate is 48–97M tokens per week. Our numbers land about 5% high, and the gap is exactly what we set aside — output tokens, which at 24 credits per multiplier unit are twelve times dearer than a cached input token and pull the achievable total down. So the derivation holds, and it confirms what the published range means: the bottom and top of every row in that table are the same workload billed at the peak and off-peak rates. Nothing about model quality, project size or your prompting varies between 48 and 97. Only the clock does.

Two smaller lines nobody has picked up

Tool calls eat plan credits. Z.ai’s credit formula has a second clause that is easy to skim past: MCP tool credit usage equals number of calls × the output multiplier. For the Web Search, Web Reader and Zread MCP servers that multiplier is 1.2, so every tool call costs 1.2 credits regardless of how much text comes back. At the implied redemption rate above that is roughly $0.0022 a call (our arithmetic). It is small, but it is a line item that no comparison of these two products mentions, and a search-heavy agent loop makes hundreds of such calls.

Web search costs identically at both vendors. Anthropic prices its server-side web search tool at $10 per 1,000 searches, which is a cent a search (our arithmetic). Z.ai’s built-in Web Search tool is listed at $0.01 per use. Same price, to the cent, arrived at independently. For a category where the two vendors are supposedly five times apart on cost, that is a striking piece of convergence — and it is the third such quiet convergence we have logged, after cache multipliers and data-residency surcharges, which we track in our Codex and Claude Code cost comparison.

The 9 September promotion, and what it does not touch

We flagged this date when we first published: GLM-5.3-Flash is currently sold at half price, and Z.ai states the promotion “ends at 24:00 on September 9, 2026 (UTC+8, Singapore time).” Re-read on that day, the pricing page still showed Flash at $0.075 input, $0.015 cached input, $0.25 output per million, against struck-through list prices of $0.15, $0.03 and $0.50. (Superseded: the discount was gone when we re-read the page on 13 September. See the 13 September update below.) GLM-5.3 itself is unchanged at $1.40 / $0.26 / $4.40 and carries no promotional marking.

Here is the part worth knowing if you are on the subscription rather than the API: the Coding Plan’s credit multipliers carry no promotional marking either. Flash is published at 2.3 / 0.56 / 8, flat, and Z.ai’s estimated token allowance table is stated in credits, not dollars. On the documentation as it stands today, 10 September changes what a Flash token costs on the API and leaves what a Flash token costs in plan credits exactly where it was. Z.ai has not published anything saying the credit multipliers will move, and we are not going to guess that they will — but subscribers should be aware that the deadline everyone is watching is an API deadline, not a plan deadline.

So what changed in the verdict?

Nothing in our recommendation, and that is worth saying plainly. GLM’s Coding Plan at $18 remains the cheapest credible way to drive Claude Code, and for a developer in the Americas the off-peak accident described above remains real money. What today’s reading changes is the confidence you should place in the comparisons you read elsewhere. If an article tells you the Lite plan gives you 400 prompts a week, it is describing a plan that was withdrawn on 30 July 2026. If it converts credits to dollars at a single exchange rate, it has assumed a linearity that Z.ai’s own multipliers do not have. And if it says GLM is five times cheaper than Anthropic on cache, it has read the API page and not the plan page.

The standing warning on this page still applies to every figure above: Claude 4.7 and later models use a tokenizer that Anthropic says produces approximately 30% more tokens for the same text, so per-token prices across these two vendors are not directly comparable, and we have not found a cross-vendor tokenisation benchmark anyone is willing to stand behind. The dollar figures are each vendor’s official list price; the ratios are our division of those figures. For the wider price landscape these two sit inside, see what AI coding assistants actually cost and our model-by-model comparison.

Sources for this update, all read 4 September 2026: Z.ai Plan Update Announcement (docs.z.ai/devpack/notice/usage-revision), Z.ai GLM Coding Plan overview (docs.z.ai/devpack/overview), Z.ai pricing (docs.z.ai/guides/overview/pricing), and Anthropic pricing (platform.claude.com/docs/en/about-claude/pricing).

Update, 6 September 2026: two Flash deadlines eleven days apart, and the discount that shrinks when you buy the plan

Since the 4 September update, Z.ai has not changed a single published price on GLM-5.3 or GLM-5.2. What has changed is that GLM-5.3-Flash now has two separate deadlines attached to it, and they are not the same offer. Most write-ups have picked up one and missed the other. Below is what each one actually is, followed by the comparison this page exists for — the one line that changes when you stop paying per token and start paying for the subscription.

Deadline one: the 50% Flash launch discount ends 9 September

Z.ai’s pricing page still shows GLM-5.3-Flash with its list prices struck through. The published note reads that the promotion “ends at 24:00 on September 9, 2026 (UTC+8, Singapore time)”. When it does, every Flash rate returns to list:

GLM-5.3-Flash, per 1M tokens	Now (promo)	From 10 September (list)
Input	$0.075	$0.15
Cached input	$0.015	$0.03
Output	$0.25	$0.50

This is an API price change. It is not, on its face, a GLM Coding Plan change — the plan bills in credits through a separate multiplier table, and that table is published on a different page. Whether the multipliers move when the API promo lapses, the official documentation does not say. We are not going to guess, and we will re-read both pages on 10 September.

Deadline two: a separate Flash usage campaign runs 3–20 September

Running underneath the price promotion is a distinct offer that applies only to paying GLM Coding Plan subscribers. Z.ai’s campaign notice states the period as 3 September to 20 September 2026, and that every day from 23:00 to 09:00 the following day (Singapore time, weekends and public holidays included), Flash usage follows special rules: through ZCode it is “Zero quota consumption” for unlimited usage, and through other supported agents the available quota is doubled. It takes effect automatically, with no activation. It applies to GLM-5.3-Flash only — select GLM-5.3 and your quota is consumed under the standard rules.

That window matters more than it looks for anyone outside Asia. Converting it (our arithmetic on the vendor’s stated Singapore times): 23:00–09:00 UTC+8 is 15:00–01:00 UTC, which lands on roughly 11am to 9pm US Eastern and 8am to 6pm US Pacific. The free-Flash window is, for a North American developer, close to the entire working day. This is the second time on this page that a Z.ai clock rule quietly favours Western users — see the peak-hours section above for the first — and it is worth saying plainly that both are accidents of timezone, not policy aimed at you, and either could be redrawn without notice.

	Flash 50% price promo	Flash usage campaign
Applies to	Pay-as-you-go API	Paid GLM Coding Plan only
What it does	Halves the per-token list price	Zero quota via ZCode; doubled quota via other agents
Hours	All hours	23:00–09:00 SGT daily
Ends	24:00, 9 September 2026	20 September 2026

The practical read: if you are evaluating Flash this week, you are evaluating it under two stacked promotions that both expire this month. Whatever number you measure before 10 September is not the number you will pay in October.

The line nobody compares: Flash is nine times cheaper on the API and exactly three times cheaper inside the plan

Here is the finding this update exists for. Everyone writing about Flash quotes the API price gap. Nobody appears to have checked whether that gap survives inside the subscription. It does not — and the difference is not small.

On the API, using Z.ai’s published list prices, Flash undercuts GLM-5.3 by roughly nine to one. Inside the GLM Coding Plan, where billing runs on the credit multipliers reproduced earlier on this page (GLM-5.3 at 6.9 / 1.7 / 24 and Flash at 2.3 / 0.56 / 8), the same swap buys you a factor of exactly three:

How much cheaper is Flash than GLM-5.3?	Input	Cached input	Output
API, at list prices	9.33x	8.67x	8.80x
API, during the current promo	18.67x	17.33x	17.60x
Inside the GLM Coding Plan	3.00x	3.04x	3.00x

The input and output rows inside the plan are 6.9 ÷ 2.3 and 24 ÷ 8 — both exactly 3.000, not approximately. So the honest answer to “should I switch to Flash to stretch my plan?” is: yes, but it buys you a third of the relief the API price list implies, and less than a sixth of what the promo price implies. If you sized your subscription tier on the assumption that dropping to Flash would cut your credit burn by an order of magnitude, the plan will cut it by two-thirds instead. That is still worth doing. It is not what the headline price comparison promises.

Note also what this does to the campaign above: during the 23:00–09:00 window, Flash through ZCode consumes no quota at all, which is not a multiple of anything. For seventeen days, the correct comparison in that window is not “three times cheaper” but “free, until it isn’t”.

Buying the subscription makes cache hits relatively more expensive, not less

This page’s original argument was that GLM-5.3’s cached-input price of $0.26 per million is higher in absolute dollars than Claude Sonnet 5’s $0.20, which inverts the usual “the Chinese model is five times cheaper” framing on cache-heavy agent workloads. Both of those numbers were re-read on 6 September and both are unchanged. What we had not checked until today is what the subscription does to that relationship. It makes it worse.

Express a cache hit as a share of the same model’s uncached input, which is the only way to compare a dollar price against a credit multiplier at all:

Cache hit, as a share of uncached input	Share
Claude Sonnet 5, Opus 5, Haiku 4.5 (published rule)	10.0%
GLM-5.3, API prices ($0.26 ÷ $1.40)	18.6%
GLM-5.3-Flash, API list prices ($0.03 ÷ $0.15)	20.0%
GLM-5.3, GLM Coding Plan multipliers (1.7 ÷ 6.9)	24.6%
GLM-5.3-Flash, GLM Coding Plan multipliers (0.56 ÷ 2.3)	24.3%

Moving GLM-5.3 from the API onto the Coding Plan takes its cache-hit share from 18.6% to 24.6% — about 33% relatively worse (our arithmetic). For Flash the same move is about 22% relatively worse. Anthropic’s cache read stays at a flat tenth of input on every model it sells, published as a multiplier rather than a per-model price, so there is nothing equivalent to drift.

The reason this matters rather than being trivia: Z.ai’s own documentation tells you the workload is cache-dominated. Its published “estimated token allowance” table is indexed by cache hit rate, and the three rows it chose to publish are 95%, 96% and 98%. The vendor is modelling a workload in which nearly all input is cache hits — which is precisely the regime where its cache pricing is least favourable relative to Anthropic’s, and the subscription widens that gap rather than closing it. None of this makes the plan a bad buy; at $18 a month it plainly is not. It does mean the cache line is the one to check before you conclude the savings are an order of magnitude.

One shape the plan flattened

A smaller structural observation, offered because it is checkable in ten seconds and we have not seen it anywhere. On the API, GLM-5.3 and Flash have different output-to-input ratios: 4.4 ÷ 1.4 is 3.14, while 0.50 ÷ 0.15 is 3.33. Inside the Coding Plan they are identical — 24 ÷ 6.9 and 8 ÷ 2.3 are both exactly 80/23, or 3.478 (our arithmetic). Anthropic, for its part, runs a flat 5.0 across Sonnet 5, Opus 5 and Haiku 4.5.

The consequence is practical: inside the plan, switching between GLM-5.3 and Flash changes the scale of your credit burn but not its shape. You cannot re-optimise a verbose-output workload by changing model tier, because output stays worth 3.478 input tokens either way. On the API you could, slightly. It is a small thing, but it is the kind of small thing that invalidates a spreadsheet.

What did not change on 6 September

We re-read all three vendors’ official pricing pages line by line today so that the comparison above rests on same-day figures rather than a mix of dates:

• Anthropic — unchanged. Sonnet 5 at $2 / $0.20 cache hit / $10; Opus 5 at $5 / $0.50 / $25; Haiku 4.5 at $1 / $0.10 / $5. Cache write multipliers still 1.25x for five minutes and 2x for one hour, cache read still 0.1x. The note that the $2/$10 introductory price “is now the standard price” and that the scheduled increase to $3/$15 “will not occur” is still on the page, as is the warning that Claude 4.7 and later use a tokenizer producing “approximately 30% more tokens for the same text”.
• Z.ai — only Flash is moving. GLM-5.3 and GLM-5.2 remain $1.4 / $0.26 / $4.4. The Coding Plan credit allowances are unchanged: Lite 2,000 per five-hour window and 10,000 weekly, Pro 12,000 and 60,000, Max 28,000 and 140,000. The off-peak half-rate rule and its Monday-to-Friday 14:00–18:00 Singapore peak window are unchanged.
• OpenAI — unchanged since 5 September. The Codex credit card still reads GPT-6 Astra at 250 / 25 / 1,250 and GPT-5.6 Sol at 100 / 10 / 500, and the per-plan message ranges are identical. Relevant here only as a control: two of the three vendors held still this week.

What we still cannot verify

• Whether the plan multipliers move when the Flash API promo lapses on 10 September. The multiplier table and the price table live on different pages and neither cross-references the other. Not stated; not guessed.
• What cached-input storage will cost. Every single row of Z.ai’s price table — every text model, every vision model, current and legacy — carries “Limited-time Free” in the cached-input-storage column. Unlike the Flash promotion, that waiver carries no published end date and no published price. It is the one line item on the entire card that you cannot budget for, and on a 95%-cache-hit workload it is not an obviously small one. Anthropic has no equivalent line: it charges a premium on the cache write and lets the cache expire, rather than metering storage.
• What one GLM credit is worth in dollars. Z.ai does not publish a conversion. We deliberately do not divide the subscription price by the weekly credit allowance to manufacture one — we made exactly that class of error on a sister page about OpenAI’s credit card and had to correct it, because a vendor’s rate card and its included allowances are two independently maintained sets of numbers that need not move together.

If you are pricing this decision across more than these two products, our wider AI coding assistant pricing comparison carries the same figures alongside the other subscription tiers, and the Claude Code alternatives page covers what else is worth testing before you commit to a year. For the OpenAI side of the same question, see our Codex versus Claude Code comparison, which tracks the credit rate card in detail.

All Z.ai, Anthropic and OpenAI figures in this section were read directly from the vendors’ own documentation on 6 September 2026. Ratios, percentages and timezone conversions are our arithmetic on those published numbers and are labelled as such.

Update, 7 September 2026: OpenAI’s rate card settles the cache argument — every text model on it prices a cache hit at exactly 10%

The central claim on this page has been narrow and awkward for the consensus: GLM-5.3 bills a cache hit at $0.26 per million tokens, and Claude Sonnet 5 bills one at $0.20. On an agent workload — the kind where the same repository context is re-sent turn after turn — that single row can outweigh the headline input price, where GLM-5.3 looks dramatically cheaper.

There is a fair objection to that argument, and until today we could not answer it. Maybe Anthropic is the outlier. Maybe a cache read priced at one tenth of fresh input is an Anthropic-specific quirk, and GLM’s ratio is unremarkable next to the rest of the industry. If so, the comparison on this page would be a coincidence dressed up as a finding.

On 7 September 2026 we read OpenAI’s published credit rate card for Codex and ChatGPT Work to find out. The answer is unusually clean: every text model on OpenAI’s rate card prices cached input at exactly 10% of ordinary input. Not approximately, not on the flagship only — on all nine text rows, with no exceptions.

OpenAI’s cached-input ratio, computed row by row

Rates below are OpenAI’s published credits per million tokens. The percentage column is our arithmetic on those official numbers.

Model	Input	Cached input	Cached as % of input (our arithmetic)
GPT-6 Astra	250	25	10.0%
GPT-5.6 Sol	100	10	10.0%
Daybreak Blue	100	10	10.0%
Daybreak Red	312.5	31.25	10.0%
GPT-5.6 Terra	50	5	10.0%
GPT-5.6 Luna	5	0.5	10.0%
GPT-5.5	125	12.50	10.0%
GPT-5.4	62.50	6.250	10.0%
GPT-5.4 mini	18.75	1.875	10.0%
GPT-Image-2 (image)	200	50	25.0%
GPT-Image-2 (text)	125	31.25	25.0%

The two image rows are the only departures, at 25.0%. Every row a coding agent would actually run sits on 10.0%.

The three-way picture

Put that next to the two vendors this page already tracks. Anthropic does not publish a percentage; it publishes a multiplier, and the multiplier is the rule itself: a cache read costs 0.1x the base input price on every model except Claude Fable 5.1 and Claude Mythos 5.1, which use 0.025x. That is an official rule, not our estimate. The GLM figures are our arithmetic on Z.ai’s published per-million prices and on the GLM Coding Plan credit multipliers.

Vendor and surface	Cache hit as % of fresh input	Source
OpenAI, all nine text models	10.0%	Our arithmetic on published credit rates
Anthropic, all models except Fable 5.1 / Mythos 5.1	10.0%	Official 0.1x multiplier
Anthropic Fable 5.1 / Mythos 5.1	2.5%	Official 0.025x multiplier
GLM-5.3, API list price	18.6%	Our arithmetic ($0.26 / $1.40)
GLM-5.3-Flash, API list price	20.0%	Our arithmetic ($0.03 / $0.15)
GLM-5.3, inside GLM Coding Plan	24.6%	Our arithmetic (1.7 / 6.9 credit multipliers)
GLM-5.3-Flash, inside GLM Coding Plan	24.3%	Our arithmetic (0.56 / 2.3 credit multipliers)

So the objection does not survive. Ten percent is not an Anthropic quirk — it is where both large American vendors independently landed, across seventeen model rows between them, with a single documented exception in each direction. GLM sits between 18.6% and 24.6% depending on whether you buy tokens or buy the subscription, which is roughly two to two and a half times the industry position on the one row that a cache-dominated agent workload leans on hardest.

The point is worth stating carefully, because it is easy to overclaim. This does not mean GLM is more expensive overall. On fresh input it is not close — $1.40 against $2.00 for Sonnet 5, and far cheaper again on Flash. What it means is that the cheapness is concentrated in the row that a well-built agent hits least often, and thinnest in the row it hits most often. Z.ai’s own documentation makes the same assumption we do: its published token-allowance estimates are built at 95%, 96% and 98% cache hit rates. The vendor and we agree about the shape of the workload; we simply followed its own assumption one row further than its pricing page does.

Two reasons you cannot turn any of this into a dollar comparison

We have deliberately compared ratios rather than prices, for two separate reasons that are each sufficient on their own.

OpenAI’s numbers are credits, not dollars. The rate card is denominated in credits per million tokens, and OpenAI states that credit purchase prices and applicable discounts depend on your plan or agreement. There is no published universal credit-to-dollar rate, so any article that converts OpenAI credits into a per-million-token dollar figure and lines it up against Anthropic or Z.ai has invented the conversion. We will not do that. Ratios inside one rate card are safe; dollars across rate cards are not.

A token is not a token across vendors. Anthropic states plainly that Claude 4.7 and later models use a newer tokenizer producing approximately 30% more tokens for the same text, with the exact increase depending on content and workload shape. A per-million-token price is a price per unit of a unit that differs between vendors. Every cross-vendor price table on the internet, including the ones on this site, should carry that warning, and ours do.

The second finding: the shape of the bill, not just its size

While we had all three rate cards open we computed one more ratio nobody seems to publish: output price divided by input price. Call it the shape of the bill. It tells you how much a verbose model costs you relative to a model that reads a lot and writes little.

Vendor and surface	Output ÷ input (our arithmetic)
GLM-5.3, API list	3.14
GLM-5.3-Flash, API list	3.33
Everything inside GLM Coding Plan	3.478 (exactly 80/23)
Anthropic, every current model	5.0, flat
OpenAI: GPT-6 Astra, GPT-5.6 Sol, Daybreak Blue	5.0
OpenAI: Daybreak Red, Terra, Luna, GPT-5.5, GPT-5.4	6.0
OpenAI: GPT-5.4 mini	6.03

Three different structures, and the practical consequence runs opposite to the headline prices. Anthropic charges a flat 5.0 everywhere, so switching Claude models changes your bill’s size but never its shape. OpenAI charges 5.0 on its two frontier models and 6.0 on everything below them — which means OpenAI’s own advice to switch to a smaller model for routine work makes the shape of your bill slightly worse at the same time as it makes the size much better. We work that trade through in detail on our Codex usage limits page. GLM is flattest of the three at roughly 3.1 to 3.5, which is the one structural respect in which it is unambiguously the friendliest of the three to output-heavy work.

If your agent reads a large repository and emits short diffs, the cache row dominates and GLM’s advantage narrows sharply. If it writes long files from short prompts, the shape row dominates and GLM’s advantage widens. Almost every comparison online reports a single multiplier as though one number answered both cases.

A cost line Anthropic charges for and OpenAI’s rate card does not mention

One more asymmetry, checked today and absent from every comparison we can find. Anthropic bills long-running agent work on two dimensions, not one: tokens at the rates above, plus session runtime at $0.08 per session-hour for Claude Managed Agents, metered only while a session is actually running rather than idle. Separately, its code execution tool gives each organisation 1,550 free hours per month and then charges $0.05 per hour per container — though code execution is free outright when used alongside web search or web fetch.

OpenAI’s Codex rate card has no equivalent hourly SKU. Cloud tasks draw from the same message allowance and the same credits as local ones; the page notes only that cloud chats may use more of your allowance than local messages. Z.ai’s Coding Plan likewise meters credits, not wall-clock time.

We are not claiming this makes Anthropic more expensive. $0.08 per session-hour is small against token spend in most sessions, and it replaces container-hour billing rather than stacking on top of it. The point is narrower: a runtime meter exists on one vendor and not the others, so a purely per-token comparison is not measuring the same set of charges on both sides. If you run agents that sit open for hours, that is a line item you should model explicitly rather than assume away.

What did not change today

We re-read both American vendors’ pricing pages in full on 7 September 2026 and confirmed the following unchanged, item by item, against our 6 September reading:

• Anthropic: Sonnet 5 at $2 / $0.20 / $10; Opus 5 at $5 / $0.50 / $25; Haiku 4.5 at $1 / $0.10 / $5; Sonnet 4.6 at $3 / $0.30 / $15; Fable 5.1 and Mythos 5.1 at $10 / $0.25 / $50. Cache write multipliers 1.25x for five minutes and 2x for one hour. The note that $2/$10 is now the standard Sonnet 5 price and that the scheduled increase to $3/$15 will not occur is still present verbatim, as is the tokenizer note.
• OpenAI: all nine text credit rows and both image rows identical to yesterday, and the per-plan local-message allowance table identical cell for cell. The note that GPT-5.6 Sol’s promotional pricing runs at least through 21 November 2026 is still present.
• Z.ai: not re-read today. The GLM-5.3-Flash 50% promotion still ends at 24:00 on 9 September 2026 Singapore time, and the Flash usage campaign still ends 20 September 2026, per our 6 September reading. Our commitment to re-read both Z.ai tables on 10 September stands — the open question is whether the Coding Plan credit multipliers move when the API promotion expires, which Z.ai has not stated and we will not guess.

One presentational change worth recording: OpenAI’s Codex pricing URL now redirects to its ChatGPT Learn documentation, and the page opens by stating that ChatGPT Work and Codex share usage, pricing, credits and limits. That is a documentation reorganisation rather than a price change, but it means Codex pricing and ChatGPT Work pricing are now formally one thing rather than two.

Still not verifiable

• What one OpenAI credit costs in dollars. Not published; stated to depend on plan or agreement. Any dollar-denominated OpenAI-versus-GLM table you see has filled this in from somewhere unofficial.
• Whether GLM Coding Plan multipliers follow the API promotion’s expiry. The two tables live on different pages and do not reference each other. Not stated; not guessed.
• Effective cache hit rates in practice. All three vendors publish prices; none publishes what hit rate real agent workloads achieve. Z.ai’s 95–98% figures are estimates in an allowance table, not measurements.

For the dollar-denominated side-by-side across every assistant we track, see our AI coding assistant pricing comparison; for why cache hit rates depend so heavily on how much context a tool re-sends each turn, see how the Claude Code context window works.

Update, 10 September 2026: Moonshot’s new flagship prices a cache hit at exactly 10% — the line everyone assumed was an American convention

Two Chinese labs published flagship prices this week, and between them they break the framing this article has been using. Moonshot’s Kimi K3 is now listed at $3.00 per million input tokens, $0.30 on a cache hit, and $15.00 output, with a 1,048,576-token context window. MiniMax’s M3 is listed at $0.30 input, $0.06 cache read, and $1.20 output under a promotion the page labels “Permanent 50% off” against a $0.60 / $0.12 / $2.40 list. Both figures are read verbatim from the vendors’ own pricing documentation on 10 September 2026.

The headline everyone else will write is that Chinese models are cheap. The number that actually matters is the one nobody lines up: Kimi K3 costs exactly 50% more than Claude Sonnet 5 on every single line — $3.00 against $2.00 on input, $0.30 against $0.20 on a cache hit, $15.00 against $10.00 on output. Three lines, one multiplier, 1.50 across the board (our arithmetic).

The 10% cache line is not an American convention — and K3 is the moment Moonshot joined it

Earlier updates to this page argued that a cache hit priced at 10% of fresh input looked like an industry floor: Anthropic documents it as a flat 0.1x rule, and every one of OpenAI’s nine text models on the Codex rate card works out to exactly 10.0% (our arithmetic). GLM-5.3’s $0.26 against $1.40 — 18.6% — stood out as the deviation. The obvious objection was that we had simply found an American habit and called it a standard.

Kimi K3 settles that. $0.30 ÷ $3.00 is exactly 10.0% (our arithmetic) — a Chinese lab, sitting precisely on the line. And it is new behaviour for that lab: Moonshot’s own previous generation is not on it. K2.6 prices a cache hit at $0.16 against $0.95, or 16.8%; K2.7 Code at $0.19 against $0.95, or 20.0% (our arithmetic, both from Moonshot’s model cards).

So the split is not China versus the US. It is a convention being adopted, and K3 is the moment Moonshot adopted it. Currently on the line: Anthropic (documented rule), OpenAI (nine models), Moonshot K3. Currently off it: GLM-5.3 at 18.6% and MiniMax M3 at 20.0% ($0.06 ÷ $0.30, and identically 20.0% at list price) — roughly 1.9x to 2.0x the convention.

“Chinese models are cheaper” stopped being a useful sentence

On the same day, from two labs in the same country: M3 input is one tenth of K3 input ($0.30 against $3.00), M3 output is one twelfth and a half of K3 output ($1.20 against $15.00), and M3 cache reads are one fifth of K3’s (our arithmetic). The spread inside the Chinese field is now larger than the spread between countries. Any sentence that treats “Chinese models” as a price tier is describing a category that no longer has a price.

Measured against Claude Sonnet 5, M3 is 85% cheaper on input and 88% cheaper on output — but only 70% cheaper on a cache hit ($0.06 against $0.20, our arithmetic). That is the same shape this article has been pointing at all along: the discount is deepest on the lines an agent loop touches least, and shallowest on the one it touches most.

Two structural facts the per-token tables hide

K3 always reasons, and it defaults to maximum effort. Moonshot’s documentation states the model “always reasons” and exposes a top-level reasoning_effort field taking low, high, or max — with max as the default. Reasoning is billed as output, and output is the $15.00 line. Out of the box, K3 is configured to spend on its most expensive column. We are not putting a number on the size of that effect, because Moonshot does not publish one.

MiniMax has a billing cliff at 512k that Kimi does not have. M3’s pricing page splits into a “≤512k input” tier and a “>512k input” tier, and crossing it exactly doubles every line: $0.30 to $0.60 input, $0.06 to $0.12 cache, $1.20 to $2.40 output. K3’s table has a single row that holds flat all the way to 1,048,576 tokens. Above 512k of input, M3’s advantage over K3 narrows from roughly 10x to roughly 5x (our arithmetic). For long-context agent work this is the difference the headline price will not tell you. Our separate write-up on how context windows actually get consumed covers why that threshold gets crossed more often than people expect.

The shape of the bill

Dividing output by input gives a rough shape of what a verbose workload does to a bill (our arithmetic). K3 lands at 5.0 — identical to Anthropic’s constant across its whole range, and to OpenAI’s Astra, Sol and Daybreak Blue. M3 is 4.0, K2.7 Code is 4.21, and GLM-5.3 remains the flattest of the group at 3.14. Flatter is friendlier to output-heavy work. Our breakdown of Codex usage limits shows the same ratio moving from 5.0 to 6.0 as you step down OpenAI’s model ladder.

One more thing worth noticing: Moonshot’s flagship refresh was a price increase. K3 against K2.7 Code is 3.16x on input and 3.75x on output (our arithmetic). The cheap Chinese coding model people are recommending is generally the previous generation, not the current flagship.

The comparability warning still applies to all of the above

Every ratio here compares list prices per token, and tokens are not a common unit across vendors. Anthropic states that its tokenizer from Claude 4.7 onward produces “approximately 30% more tokens for the same text.” A per-token quote from Anthropic and a per-token quote from Moonshot are therefore not measuring the same quantity of work, and we have deliberately not compounded that 30% into any of the multipliers above. Treat the cache-ratio finding — which is internal to each vendor’s own table and so immune to the tokenizer problem — as the sturdiest number on this page. See also our full pricing comparison across coding assistants.

What the vendors have not published

Not disclosed as of 10 September 2026, and therefore not stated here: K3’s release date; whether K3 is eligible for any batch discount (the pricing page lists a single row with no batch tier); the retirement status of K2.5 and Moonshot V1; M3’s Token Plan monthly fee, release date, and total context ceiling beyond the 512k billing boundary. Separately, Moonshot’s K3 page carries an official warning that its web_search tool is mid-update and is not recommended for near-term use, and Moonshot’s API platform has moved from platform.moonshot.ai to platform.kimi.ai. Finally, Z.ai’s GLM-5.3-Flash promotion was documented to end at 24:00 on 9 September 2026 (UTC+8), but as of this update the official page still displays the discounted figures and the original end-date note — so the GLM Flash numbers elsewhere on this page are left unchanged rather than guessed at.

Update, 13 September 2026: the Flash discount is gone — and the date Z.ai published for it was not the date it ended

Re-reading docs.z.ai/guides/overview/pricing on 13 September 2026, the GLM-5.3-Flash row has changed. There is no strikethrough, no discounted figure and no expiry tip. The table simply lists $0.15 input, $0.03 cached input, cache storage “Limited-time Free”, $0.50 output, per million tokens. Those are the list prices we had been printing in brackets since 30 August. They are now the prices.

Every other row on that page is unchanged from our 12 September reading: GLM-5.3, GLM-5.2 and GLM-5.1 at $1.40 / $0.26 / $4.40; GLM-5 at $1.00 / $0.20 / $3.20; GLM-4.7 at $0.60 / $0.11 / $2.20; GLM-4.7-FlashX at $0.07 / $0.01 / $0.40; GLM-4.7-Flash and GLM-4.5-Flash free; web search at $0.01 per call. The withdrawal is confined to one model.

The dates, and why the published expiry was not usable

This is the part we would not know if we were not reading the page every day, and it is worth more than the price change itself.

Z.ai published a specific end for this promotion. The tip on the pricing page said it ended at 24:00 on 9 September 2026, Singapore time (UTC+8). We quoted that date in this article and in our AI coding assistant pricing comparison, and we told readers their Flash costs would double from 10 September.

That is not what happened. On 10 September we re-read the same page and the discount was still live: struck-through list prices, discounted prices beside them, and the same expired tip still sitting underneath. The page did not change on the date the page itself gave. When we next got a clean read of it, on 13 September, the discount was gone. We did not retrieve the page on 11 or 12 September, so the honest statement of when it changed is some time between 11 and 13 September 2026. Z.ai has not posted a withdrawal date, and we are not going to invent one on the vendor’s behalf.

The general lesson is small but it costs real money: a vendor’s published promotion end date tells you when the vendor intends to stop discounting, not when your invoice changes. The invoice changes when the price table changes, and on this occasion the two were at least a day apart and possibly four. If you are building a cost model around a promotional rate, the date to watch is the one on which the table stops showing it, which means somebody has to look.

The level doubled. The shape did not move at all.

The obvious sentence to write is “GLM-5.3-Flash just got 100% more expensive”, and it is true: $0.075 to $0.15, $0.015 to $0.03, $0.25 to $0.50 — each line exactly doubled (our arithmetic on Z.ai’s published figures). But it is the least useful true sentence available, because the discount had been applied uniformly to all three lines. Halving everything and then unhalving everything leaves the internal proportions of the bill untouched.

Two ratios we track on this site make the point. The output-to-input ratio — what we have been calling the shape of the bill — was $0.25 ÷ $0.075 = 3.33 under the discount and is $0.50 ÷ $0.15 = 3.33 today (our arithmetic). The cached-input share was $0.015 ÷ $0.075 = 20.0% and is $0.03 ÷ $0.15 = 20.0% today (our arithmetic). Neither number moved by a hundredth.

That has a practical consequence that we have not seen anyone state. If you picked Flash because of the shape of its bill, the withdrawal changes nothing about that decision. Flash is still the flattest-billing model in Z.ai’s line-up relative to GLM-5.3 at 3.14, still cheaper on cached input as a proportion of input than GLM-5.3’s 18.6%, and still on the far side of the roughly 10% cache convention that Anthropic, OpenAI and now Moonshot all sit on (see the framework earlier in this article). Every shape-based conclusion in this article survives the price change intact. Only the level-based ones need redoing — and those are the ones written as absolute dollars per task.

What it does to the comparison against Claude Code

Agent workloads are cache-dominated: the same repository context is resent on every turn, so most input tokens are cache hits rather than fresh input. The number that actually drives an agent bill is therefore not the headline input price but the blend of cache-hit price and fresh-input price. Below is that blend at a 95% cache-hit rate, which is a common figure for a long coding session. The dollar figures in the first two columns are vendor list prices; the blended column is our arithmetic (0.95 × cache hit + 0.05 × input).

Model	Input / 1M	Cache hit / 1M	Blended at 95% hits (our arithmetic)
Claude Sonnet 5	$2.00	$0.20	$0.29
GLM-5.3	$1.40	$0.26	$0.32
GLM-5.3-Flash (list, 13 Sep 2026)	$0.15	$0.03	$0.036
GLM-5.3-Flash (withdrawn discount)	$0.075	$0.015	$0.018

Read the first two rows before the last two. On a cache-heavy agent load, Claude Sonnet 5 blends out below GLM-5.3 — $0.29 against $0.32 — because Anthropic’s cache-hit price of $0.20 undercuts Z.ai’s $0.26 even though Anthropic’s headline input price is 43% higher (our arithmetic). That reversal is the single most useful line on this page and the Flash withdrawal does not touch it, because it is a statement about GLM-5.3, not about Flash.

What the withdrawal does touch is the size of the Flash advantage. Flash blended at roughly $0.018 per million under the discount and roughly $0.036 today, against Sonnet 5’s $0.29 — so the ratio moves from about 16× cheaper to about 8× cheaper (our arithmetic). Halving an advantage of that size does not change which model is cheaper. Anyone who chose Flash on price alone has no reason to move, and anyone who was deciding between Flash and Sonnet 5 on a spreadsheet where the margin was under 2× was not really deciding on price.

The caveat that applies to every row of that table. Anthropic states that Claude 4.7 and later use a new tokenizer which produces roughly 30% more tokens for the same text. Per-token list prices from different vendors are therefore not directly comparable, and we have deliberately not folded that 30% into any of the multiples above. Treat the cross-vendor comparisons as directional and the within-vendor ones (Flash against GLM-5.3, discount against list) as exact.

Two things this update does not change. The GLM Coding Plan’s credit multipliers still carry no promotional marking, so subscription users were never on the discount and see nothing today. And the long-context billing question is separate from all of it — how a vendor charges once a request gets long is a different mechanism from what it charges per token, and we cover it in the Claude Code context window piece. If you want the per-task view rather than the per-million view, our model comparison converts these rates into representative task costs.

Update, 16 September 2026: a second Chinese vendor now advertises 50% off-peak — and still will not say when off-peak is

A large part of the comparison above turns on one number: the off-peak rate. Z.ai documents that during off-peak hours usage is charged at 50%, and our cost model leans on that discount heavily. On 10 September 2026 DeepSeek repriced its API alongside the launch of V4.1-Flash, and attached the identical structure to it: off-peak rates are 50% of peak rates, with an explicit recommendation to schedule flexible workloads off-peak.

Two vendors, independently, landed on exactly the same discount ratio. That convergence is mildly interesting on its own. What is more interesting is the thing both of them leave out. In neither case does the vendor documentation state the peak and off-peak windows in text. We could not find the hours for DeepSeek in prose on the pricing or release pages, and the discount is advertised without them.

This matters for our own arithmetic above, so we should be explicit about it rather than let it sit in a footnote. Our estimate assumed a developer whose working hours fall almost entirely inside the off-peak window, which is what makes the headline saving so large. That assumption is doing more work than any price in this article, and it is the one number here that you cannot verify from vendor documentation. If your team is distributed, or your agent runs on a CI trigger rather than on your keyboard, the off-peak share is an empirical question about your own traffic, not a property of the plan. We would rather flag the soft spot in our own model than present a discounted figure as though it were a list price.

The cache-hit line, now conceded by a vendor

When we first published this comparison we argued that the cache-hit rate, not the headline input price, is the line that decides what an agent workload actually costs, and that almost nobody compares it. We were making an inference from the pricing tables. DeepSeek has now made the same argument in its own marketing.

Its V4.1-Flash release notes lead on cache compression rather than on token price, and say in as many words that cache-hit charges often account for a large share of agent costs, and that compressing the cache cuts those costs significantly. The vendor claims the new architecture needs roughly a quarter of the HBM and an eighth of the SSD storage of the previous generation for its KV cache.

Read that as a competitive signal rather than as a spec. When a vendor chooses to advertise the cache line instead of the per-token line, it is telling you which line its customers are actually complaining about. The headline price comparison that every article runs — input and output per million tokens, side by side — is being quietly abandoned by the vendors themselves in favour of the number underneath it. That is the same conclusion we reached on this page by a different route, and it is why the cache-hit row in the table above is the one to read first.

One limit on all of this: we are not quoting DeepSeek per-token or cache-hit prices, because the V4.1-Flash price list was published as an image rather than as text, and we do not transcribe pictures into price tables. The retirement of V4-Flash and the same-day reversal on V4-Pro are covered in our AI coding assistant pricing comparison, and the vendor-claimed benchmark numbers are in our coding model benchmark comparison. For how context-length thresholds interact with cached workloads, see the Claude Code context window guide.

Update, 17 September 2026: Z.ai’s own notice page says weekends are off-peak all day — which settles two of the seven days we had been guessing about

Yesterday we added a section to this page arguing that the single number carrying the most weight in this comparison was not a price at all. It was our assumption about when off-peak is. We said plainly that it was the one figure we could not verify against vendor documentation. Today we can verify part of it, so we are correcting ourselves in public.

Z.ai publishes a Plan Update Announcement at docs.z.ai/devpack/notice/usage-revision, dated 30 July 2026 on the page itself. We read it on 17 September 2026 — 49 days after it was published. It contains this sentence:

“Existing active subscriptions are not affected. In addition, usage on weekends will be deducted at off-peak rates all day.”

That is the first time we have seen this vendor state an off-peak period in text rather than leaving it to be inferred from a peak window. It is worth being precise about how much it actually settles.

Two days official, five days still inferred

Our claim earlier on this page — that a North American developer working normal hours is off-peak essentially all of the time — rested entirely on the peak window (Monday to Friday, 14:00–18:00 UTC+8) landing in the middle of the North American night. That inference is unchanged. What changed is that Saturday and Sunday are now off-peak by official statement rather than by our arithmetic, for every subscriber in every timezone.

So the honest scoring is: two of seven days confirmed from vendor text, five of seven still resting on our reading of a peak window that the pricing page does not restate. We are not going to describe that as verified. It is partially verified, and the unverified part is still the load-bearing part for anyone outside Asia-Pacific.

Counting hours rather than days (our arithmetic, from the official peak window): a four-hour peak window on five weekdays is 20 hours. A week is 168 hours. That leaves 148 of 168 hours — 88.1% of the week — outside the peak window, of which the 48 weekend hours are now officially off-peak. One caveat we cannot close: the announcement does not say which timezone defines a “weekend.” The peak window is defined in UTC+8, so UTC+8 is the reasonable assumption, but the vendor does not state it and we are not going to invent it. If you are in the Americas, your Friday evening and Sunday evening sit near that boundary.

The bigger find: the off-peak discount is not one number, it is three — and which one you get depends on when you signed up

The same announcement makes something explicit that no comparison table we have seen accounts for. This vendor runs two different multiplier systems side by side, and the peak penalty is not the same in both.

Plan vintage / model	Off-peak multiplier	Peak multiplier	Cheapest-to-dearest spread
Current credits system (all models)	0.5× (official wording: “charged at 50% of the standard credit rate”)	1× standard	2×
Legacy Plan V2 — GLM-5.3	1×	3×	3×
Legacy Plan V2 — GLM-5.3-Flash	0.4×	1.2×	3×

Read the spread column. On the current credits system the worst hour of the week costs twice the best hour. On Legacy Plan V2 the worst hour costs three times the best hour. Two subscribers running identical workloads on identical models, differing only in which month they subscribed, face different penalties for working at the same time of day.

This is the structural reason these plans are hard to put in a comparison table, and we think it is worth naming. A pricing table has one cell per model. This vendor’s actual price has one value per (model × plan vintage × hour of day × day of week). A table with one cell per model cannot represent that without picking one of those values and silently dropping the rest — which is what every table that quotes a single number for this plan is doing, including, until today, ours. We would rather show you the shape of the thing than flatten it.

What we would actually do with this

If you are on Legacy Plan V2, moving heavy agent runs to the weekend is worth three times more than the same move is worth on the current credits system. Long refactors, whole-repository review passes, batch migrations, anything you kick off and walk away from — that is the work that moves, and the weekend line means it moves without you having to be confident about any timezone inference at all. Interactive work does not move, because you do not choose when you are typing.

And a process point that cost us 49 days: for this vendor, the pricing page is not the source of truth. The sentence that changes the arithmetic for every subscriber on earth was sitting on a notice page that the pricing page does not link to. If you are building a budget against a vendor like this, put its /notice or changelog directory on your reading list, not just its pricing page. We had been re-reading the pricing table daily and confirming, correctly, that nothing in it had changed — while the thing that mattered was published elsewhere and never touched the table.

One standing warning applies to everything above, as it does everywhere on this site: per-token and per-credit prices are not comparable across vendors, because tokenizers differ. Anthropic states that Claude 4.7 and later use a different tokenizer that produces roughly 30% more tokens for the same text. A rate card comparison that ignores that is comparing two different units. We keep the fuller version of that argument, with the cross-vendor rate tables, on our AI coding assistant pricing breakdown.

For how these models compare on capability rather than billing mechanics, see our comparison of AI coding models. And if the decision you are actually making is between the two Western incumbents rather than a Chinese challenger, our Codex vs Claude Code comparison is the page to read instead.

Sources for this update: Z.ai Plan Update Announcement, docs.z.ai/devpack/notice/usage-revision (page-stated publication date 30 July 2026), read 17 September 2026; Z.ai API pricing page, read 17 September 2026, unchanged from our 14 September reading. Nothing in this section is derived from third-party reporting. Where the vendor does not state something — the timezone governing “weekends”, and the peak window’s absence from the pricing page — we say so rather than filling the gap.

Update, 19 September 2026: both of the time-boxed offers on this page are now over or end tomorrow, and here is what that does to the comparison

This section exists to keep the page honest as two dates pass, and to retire a commitment we made earlier in it and did not keep.

First, the retraction. An earlier update on this page says our commitment to re-read both Z.ai tables on 10 September stands. We did not re-read them on 10 September. That sentence has been sitting here for nine days describing a check that did not happen, which is worse than not having promised it. Our most recent reading of Z.ai’s published tables was 17 September 2026, and everything below is stated against that reading rather than against today.

Second, the two dates. Both offers described above are quoted verbatim from Z.ai’s own notices, and both are dated:

• The GLM-5.3-Flash 50% price promotion ended at 24:00 on 9 September 2026, Singapore time. As of today that is ten days in the past, and every figure on this page that was quoted at the promotional rate describes a price you can no longer get.
• The Flash usage campaign — paid GLM Coding Plan subscribers only, 23:00–09:00 SGT daily, zero quota consumption via ZCode and doubled quota via other agents — runs through 20 September 2026, which is tomorrow.

Whether Z.ai extends, replaces or simply lets either offer lapse is officially not published as of our 17 September reading, and we are not going to guess at it. What we can do is say precisely what happens to the comparison if nothing is announced, because that part is arithmetic rather than prediction.

A 50% discount means the number you compared was exactly half

This sounds too obvious to write down, and it is the single most common mistake we see in vendor selection. A 50% promotion does not make a model “a bit cheaper” for the duration. It means that the post-promotion price is exactly 2.00x the price you evaluated, and that the doubling arrives on a date the vendor already told you.

Doubling one column is more than enough to flip most of the comparisons on this site. We showed on 18 September that Gemini Flash’s entire 62.5% price advantage over Claude Sonnet 5 becomes a 25% advantage the moment Google’s announced 1 January 2027 doubling lands — the same multiplier, a different vendor, an identical mechanism. A time-boxed price is not a price. It is a price and a date, and a comparison that carries the first without the second has a defect with a known expiry.

The practical test, which costs nothing: before choosing a vendor on a promotional rate, redo the comparison at list price and see whether your decision changes. If it does not, the promotion is a bonus and you can stop worrying. If it does, you have not chosen a vendor — you have chosen a date on which you will have to choose again, and you should know which date that is before you build on it.

The quota campaign fails in the opposite direction, and that is why it is easier to miss

The second offer is not a discount at all. It changes quota consumption, not per-token price: zero quota via ZCode and doubled quota via other agents, during a nightly window. The two offers therefore expire in two completely different ways, and only one of them shows up on an invoice.

• When a price promotion ends, your bill rises and your throughput is unchanged. You find out from finance, quickly, and the size of the surprise is exactly the discount.
• When a quota campaign ends, your bill is unchanged and your throughput falls. Nothing anomalous appears anywhere in your cost reporting. What you observe is that overnight jobs that used to finish stop finishing, and the natural response — upgrade the plan — raises the bill days or weeks later, in a line item that looks like a decision you made rather than an offer that expired.

The second one is more dangerous precisely because it is cheaper on the day it happens. A price increase announces itself. A quota reduction is indistinguishable from your workload having grown, and by the time you have diagnosed it you have usually already paid to fix it.

This is the third distinct species of vendor “free” we have had to separate on this site in three weeks. On 18 September we distinguished a standing allowance (a number that renews, like Anthropic’s 1,550 free container-hours per month) from a preview (a zero price with no end date published, like Google’s Antigravity sandbox), with the rule that allowance is followed by a quantity and preview is followed by a tense. A dated campaign is the third kind, and it is the most tractable of the three: it has both a quantity and an end date, which means there is no excuse for being surprised by it, only for not having written the date down.

We will re-read Z.ai’s tables after 20 September and record what actually happens rather than what we expect. Given the preceding paragraph about commitments we did not keep, treat that as a statement of intent and check the date on the next update.

For how the same expiry problem looks when the vendor publishes the increase more than a hundred days ahead, see our cross-vendor pricing breakdown. For surcharges that behave like a permanent version of this — triggered by prompt size rather than by a calendar — see our context-window pricing analysis. And if a compliance requirement is going to constrain which vendors are even eligible, the residency surcharges are covered in Codex vs Claude Code.

Update, 20 September 2026: Z.ai now publishes a weekly token range for each plan, and the range is exactly 2.00x wide because of what time of day you work

Two things happened on 20 September 2026 that change what this page can tell you. The first is that we finally kept a promise this page made and then broke. The second is that Z.ai’s Coding Plan documentation carries an official table we had not recorded before, and it answers the question this whole comparison exists to answer: for a fixed monthly subscription, how many tokens do you actually get?

First, the promise

On 19 September we wrote here that we had failed to re-read Z.ai’s published tables on 10 September after saying we would, and we committed in writing to re-read them after 20 September and record what actually happened. We read the pages verbatim on the morning of 20 September 2026. Here is what we found, including the part that is unsatisfying.

The GLM-5.3-Flash usage campaign page is still live on its own final day. The campaign period is stated verbatim as “Campaign period: September 3, 2026 to September 20, 2026”, and both the Tip banner at the top of the Coding Plan overview and the Tip below the token allowance section are still advertising it. So the honest report is: on the last day of the offer, nothing had been withdrawn yet. Whether the page comes down on 21 September, whether the campaign is extended, or whether it is replaced with a successor offer is officially not published, and we are not going to guess. We will read the page again after the date has passed and record the outcome here, the same way we are recording this.

That is a smaller finding than it looks, but it is the reason we bothered: the useful habit is not “assume the vendor executes on the announced date”, it is “read the page on the day after the date”. We have now been wrong in both directions on this site — once assuming an offer was still running when it had ended, once assuming an offer had ended when the page was still up. The date on the page tells you when to look, not what you will see.

Second, the table: Z.ai publishes an “Estimated Token Allowance” in millions of tokens per week

This is the number that every “is the GLM Coding Plan cheaper than Claude Code” argument on the internet has been estimating, including, until today, ours. Z.ai publishes it directly. The figures below are the vendor’s own published estimates, read verbatim on 20 September 2026, in millions of tokens per week:

Cache hit rate	Model	Lite	Pro	Max
95%	GLM-5.3	48–97	290–580	676–1,352
95%	GLM-5.3-Flash	146–292	877–1,755	2,047–4,095
96%	GLM-5.3	50–99	297–595	694–1,387
96%	GLM-5.3-Flash	150–300	900–1,801	2,101–4,202
98%	GLM-5.3	52–104	313–627	731–1,463
98%	GLM-5.3-Flash	158–317	950–1,900	2,217–4,433

The range is not uncertainty. Z.ai states the two ends explicitly: the top of each range assumes “All usage occurs during off-peak hours and is charged at 0.5x the standard credit rate”, and the bottom assumes everything runs at the standard peak rate. That single sentence is what makes the rest of this section possible, because it means the width of every range in the table has a known cause.

Every row is exactly 2.00x wide, and that is the second-biggest lever on the page

Check any row. 290 to 580. 676 to 1,352. 877 to 1,755. 2,217 to 4,433. Every single range in the official table has a top that is exactly twice its bottom, give or take the rounding in the last digit (our arithmetic). This is not a coincidence and it is not a confidence interval. It is the 0.5x off-peak multiplier expressed as a range, which means:

What you bought is not a quantity of tokens. It is a quantity that varies by 2.00x depending on what hour of the day your jobs run, and the vendor has told you so in a table most readers will skim as “roughly 300 to 600 million, call it 450”. Calling it 450 is the error. There is no workload that lands in the middle by accident; batch jobs land at the top, interactive sessions during Chinese business hours land at the bottom, and the midpoint describes almost nobody.

Ranking the three levers, which is the part nobody writes

The table contains three independent variables, and once you compute their sizes they turn out to be nothing like equally important (all of the following is our arithmetic on Z.ai’s published figures):

1. Model choice: about 3.04x. On the same plan at the same cache hit rate, GLM-5.3-Flash returns roughly three times the weekly tokens of GLM-5.3. Lite at 95%: 146 versus 48, a factor of 3.04. Pro: 877 versus 290, a factor of 3.02. Max: 2,047 versus 676, a factor of 3.03. The ratio is stable across all three plans and all three cache-hit rows.
2. Scheduling: exactly 2.00x. Peak versus off-peak, as described above.
3. Cache hit rate: about 8%. Moving from a 95% cache hit rate to a 98% cache hit rate — three percentage points, which in practice is real engineering work on prompt prefix stability — raises the Lite GLM-5.3 floor from 48 to 52, a gain of 8.3%. The same move on Pro is 290 to 313, a gain of 7.9%. On Max, 676 to 731, a gain of 8.1%.

Put that ranking next to where the industry actually spends its attention. Prompt caching is the optimisation everyone writes tutorials about, and on this table it is worth roughly 8%. Moving your jobs to off-peak hours is worth 100% — about twelve times as much — and requires no engineering at all, only a scheduler (our arithmetic). Choosing the smaller model is worth more still.

We want to be careful here, because this is a comparison of levers, not advice to pull them. Flash is a different model and the tokens are not interchangeable with GLM-5.3’s; a 3.04x allowance advantage is worth nothing if the smaller model needs four attempts. Cache hit rate improvements also reduce latency, which this table does not price. The point is narrower and it is about attention: the cheapest lever on this page is the one that shows up in no benchmark and no blog post, because it is a cron expression rather than a technique.

The corner-to-corner spread on one identical subscription is about 6.6x

Combine the levers and the table stops looking like a price list. On Lite, the worst corner — GLM-5.3, entirely during peak hours, at a 95% cache hit rate — is 48M tokens per week. The best corner of the same Lite subscription — Flash, entirely off-peak, at 98% — is 317M. That is 6.60x (our arithmetic). On Pro the same two corners are 290 and 1,900, or 6.55x. On Max, 676 and 4,433, or 6.56x.

None of those corners is a different plan. Nobody upgraded. The subscription line on the invoice is identical at both ends of a 6.6x difference in what you are allowed to run, and the entire difference is decided by three choices the purchase page does not ask you to make.

For a developer outside China, the 2.00x scheduling lever is not a decision at all

Here is the part that made us think this table is worth a section rather than a line. Z.ai’s peak window is Monday to Friday, 14:00 to 18:00, UTC+8. Weekends are off-peak all day.

That is four hours on five days: 20 hours out of the 168 hours in a week, or 11.9% of all wall-clock time. The other 88.1% is already discounted (our arithmetic). A range whose bottom end requires you to do all of your work inside one-eighth of the week is a range whose bottom end almost nobody will ever see.

Now convert the window. 14:00–18:00 UTC+8 is 23:00–03:00 US Pacific on the previous day, and 02:00–06:00 US Eastern (our arithmetic). A developer on the US West Coast working ordinary hours cannot reach Z.ai’s peak window at all without deliberately staying up past eleven at night. For that reader, the published range is not a range: the plan delivers the top number, and the bottom number describes a schedule they would have to work at to achieve.

This cuts the other way too, and it is the reason we are stating it as geography rather than as a tip. A team in Shenzhen or Singapore doing interactive coding work in the afternoon sits squarely in the peak window during exactly the hours humans want to work. The same subscription, at the same price, is worth up to twice as much to the buyer who is further from the vendor’s own timezone. We have not seen that stated anywhere, and it follows directly from two published facts: a discount window defined in UTC+8, and a 0.5x multiplier.

What we are not going to tell you

Z.ai’s documentation accompanies this table with a claim we are flagging rather than repeating as fact: “By fully utilizing the off-peak discounts, you can save up to 92% compared with pay-as-you-go calls to the GLM-5.3 standard API.” That is a vendor claim. We are not able to check it, for a reason worth being explicit about: the comparison base is not specified in enough detail to reconstruct — we do not know which cache hit rate, which model, or which plan the 92% is computed against, and “up to” is doing load-bearing work in that sentence. What we can verify is the internal arithmetic of the table itself, which is the 2.00x, 3.04x and 8% above, and those we have shown our working for.

We also still do not convert Z.ai credits into dollars anywhere on this site, and this table does not change that. It is denominated in tokens per week, which is the unit your logs are already in, so it does not need converting.

Finally, the standing warning that applies to every cross-vendor comparison on this site applies to this one with particular force. Anthropic states that Claude 4.7 and later use a different tokenizer that produces approximately 30% more tokens for the same text. A weekly allowance denominated in tokens is therefore not directly comparable across vendors: 300 million Z.ai tokens and 300 million Anthropic tokens are not 300 million of the same thing. The allowance table above tells you what you can run on Z.ai. It does not, by itself, tell you how that compares to a Claude subscription, and any article that divides one by the other is quietly assuming the tokenizers match.

Related reading on this site: our full cross-vendor breakdown in AI coding assistant pricing, the subscription-side view in Claude Pro usage limits, and the endpoint-level surcharges that behave the same way in Codex vs Claude Code.

Update, 21 September 2026: the promotion ended the way promotions actually end, which is not on the date and not all at once

We put a promise in writing twice on this page. On 19 September we said we would re-read the Z.ai documentation after the stated end date and record whatever we found. On 20 September we did, and the honest answer was that nothing had happened: the GLM-5.3-Flash Usage Campaign was still being advertised on its own final day, two promotional tips still live on the DevPack overview page. We said we would read it again. Here is the result, and this closes the loop.

On 21 September 2026, both promotional tips are gone. The tip at the top of the DevPack overview page and the tip below the Estimated Token Allowance table have both been removed. The campaign notice page itself is still reachable and its text is unchanged, still reading that the campaign period ran from 3 September 2026 to 20 September 2026. So the campaign is over, the advertising is down, and the announcement remains as a historical record.

What two consecutive readings taught us that one reading could not

If we had only read the page on the final day, we would have concluded the promotion was being extended. If we had only read it the day after, we would have concluded it ended cleanly on schedule. Neither conclusion would have been right, and both would have been stated with confidence. The thing that was actually true is only visible from two readings: the banners came down after the end date, not on it.

This upgrades the rule we wrote yesterday. Yesterday we said a date tells you when to look, not what you will see. Today we can be more specific. A promotion has at least three separate clocks: the stated end date, the moment the advertising is withdrawn, and the moment the announcement page is taken down or archived. Vendors routinely run these on different schedules, and in this case the third clock has not moved at all. The practical consequence is blunt: the presence of a campaign page is not evidence that the campaign is running, and the absence of a banner is a better signal than either.

We have now been wrong about this in both directions on the same promotion within one month. Earlier in September we assumed a promotion was still live when it had already closed. On 20 September we assumed the banners would be down on the end date and they were not. Writing both of those down is more useful to a reader than a page that only ever records the times we happened to be right.

Nobody is going to email you about this

The teams most exposed to this are the ones who did the sensible thing and built the promotional terms into a cost model. During the campaign, ZCode usage was described by Z.ai as zero quota consumption and other agents as having available quota doubled, in a daily window from 23:00 to 09:00, with ZCode version 3.10 or later required. As of 21 September none of those terms apply. There was no notification. The only signal was a banner that stopped appearing on a documentation page you were not watching.

The operational rule that follows is dull and works: put the stated end date in a calendar, not in your memory of what the page said, and re-read the page at least twice after that date. Two readings, because as shown above one reading of this particular change would have produced a confident wrong answer either way.

What the list prices are now that the promotion is not in the way

These are vendor list prices read verbatim on 21 September 2026, per million tokens, in the order input / cached input / output. GLM-5.3-Flash is $0.15 / $0.03 / $0.50. GLM-5.3 is $1.4 / $0.26 / $4.4. GLM-5.2 is listed at the same numbers as GLM-5.3. GLM-5 is $1 / $0.2 / $3.2. Cached Input Storage is marked Limited-time Free across the line, and Web Search is $0.01 per use.

One line is new to our records rather than new to the world: GLM-5.3-FlashX at $0.37 / $0.075 / $1.25. We have not previously listed this SKU, which is not the same thing as it being newly introduced. Whether GLM-5.3-FlashX is a recent addition is officially not published, and we are not going to infer a launch date from our own filing gaps. We have written up what the three-rung price ladder implies on our Claude Code alternatives page, because the interesting part of it is a choice between rungs rather than a choice between vendors.

A test of whether yesterday was worth writing

Yesterday this page argued that a subscription plan has three cost levers and that they rank in an order opposite to the attention they get: model choice at roughly 3.04x, scheduling at exactly 2.00x, and cache hit rate at roughly 8%. A fair question to ask today is whether the campaign ending invalidates any of that.

It does not, and the reason is worth naming, because it is a reusable test. Every one of those three numbers came from the standing multiplier table and the peak-hours window, not from the promotion. The off-peak rate is still 0.5x. The peak window is still Monday to Friday, 14:00 to 18:00 Singapore Standard Time. None of that moved when the banners came down. An argument built on the permanent structure of a price list survives the promotion that was sitting on top of it; an argument built on promotional terms expires with them, usually without telling you. If you want a quick way to judge whether a pricing article is worth keeping in your notes, check whether its conclusions would still hold after the current discount ends.

Closing this thread

The campaign is over, the advertising is withdrawn, and the notice page stands as a record. This page is no longer tracking it, and we are deliberately not making a third promise about it. If Z.ai runs a successor campaign we will write that up as its own event. What carries forward is not the promotion but the method: read the page, read it again, and write down the times the second reading contradicted the first.

For how this interacts with what Anthropic charges for the same work, see our AI coding assistant pricing breakdown, which today gained a section on the part of caching that neither vendor puts in the headline table. If you are comparing subscription allowances rather than API rates, our Claude Pro usage limits page explains why the two units cannot be divided into each other.

Prices and campaign terms in this section were read verbatim from vendor documentation on 21 September 2026. Ratios and break-even figures elsewhere on this page are our arithmetic on those published numbers. Token counts are not comparable across vendors: Anthropic states that its tokenizer produces approximately 30% more tokens for the same text, so per-token prices from different vendors cannot be divided into each other without adjusting for that.

Update, 22 September 2026: a third subscription now sells the same $18 shape, and the one number you would need to compare them is the one nobody publishes

Everything above compares a Chinese coding subscription against Anthropic’s. As of today there is a third product with almost exactly the same shape, and it changes the question this page is asking. Alibaba sells a Model Studio Token Plan for Individual: a monthly subscription, metered in credits, with a five-hour bucket and a seven-day bucket, sold through an Anthropic-compatible endpoint that the vendor explicitly documents as working with Claude Code. Its middle tier is $18 a month. The GLM Coding Plan discussed above starts at 18 USD. That is not a coincidence of rounding; it is two vendors independently deciding what the entry price for a serious coding subscription is.

One disclosure first, because it is the honest way to report this. The vendor blog announcing this plan is dated 19 July 2026. We read it on 22 September 2026. That is roughly a two-month gap between the product existing and it reaching our records (our arithmetic). We are reporting a gap in our own monitoring, not a launch date, and nothing below should be read as “new this week.”

What Alibaba actually publishes

Three tiers, each with two credit buckets (per five hours, and per seven days), a list price, and a discounted early-bird monthly price. These are the vendor’s official list figures, read from the announcement blog on 22 September 2026:

• Lite — 700 credits per 5 hours, 2,500 credits per 7 days. $8 list, $6 early-bird monthly. 1 to 2 concurrent agents.
• Standard — 3,000 credits per 5 hours, 10,000 per 7 days. $25 list, $18 early-bird monthly. 3 to 4 concurrent agents.
• Pro — 12,000 credits per 5 hours, 40,000 per 7 days. $80 list, $68 early-bird monthly. 6 to 8 concurrent agents.

Longer commitments are also published: quarterly at $17 / $51 / $195, and yearly at $65 / $195 / $770. The vendor also claims “approximately 3x more usage” than pay-as-you-go, which we record as a vendor claim and do not use in any calculation below.

The five-hour window is not the binding limit. The weekly one is, by roughly ten to one.

Every subscription of this shape advertises the short window, because the short window is the one that sounds generous. Do the division. A seven-day period contains 33.6 five-hour windows (168 hours divided by 5; our arithmetic). If you could actually spend a full five-hour bucket in every one of them, the Lite tier would deliver 700 x 33.6 = 23,520 credits a week. The published weekly ceiling is 2,500 — about 10.6% of it (our arithmetic).

Standard and Pro land in the same place: 3,000 x 33.6 = 100,800 against a 10,000 ceiling, and 12,000 x 33.6 = 403,200 against 40,000 — 9.92% in both cases (our arithmetic). The design is consistent across the ladder, and it says something specific: the five-hour bucket is sized so that you can burst, and the weekly bucket is sized so that you cannot sustain it. You are being sold the burst and billed for the sustain.

The practical consequence is that a five-hour allowance tells you nothing useful about whether a plan fits your week. Divide the weekly number by seven and compare that against a normal working day. On Standard that is about 1,429 credits a day (our arithmetic) — and the five-hour figure of 3,000 is more than twice that, which is to say you can exhaust two days of budget in one afternoon without ever touching the limit the marketing page leads with.

The per-credit ladder flattens almost immediately

Price per weekly credit, at the early-bird monthly rates (our arithmetic): Lite $6 / 2,500 = $0.00240; Standard $18 / 10,000 = $0.00180; Pro $68 / 40,000 = $0.00170. Stepping from Lite to Standard buys you 25% off per credit. Stepping from Standard to Pro buys you 5.6% (our arithmetic).

That shape matters when you are choosing a tier. The volume discount is essentially all spent on the first step; after that you are buying headroom and concurrent agents, not cheaper tokens. So the correct reason to move from Standard to Pro is “I need 6 to 8 agents running at once” or “I keep hitting the weekly wall,” never “it will be cheaper per unit of work.” It will be 5.6% cheaper per unit of work, which is inside the noise of almost any real workload.

The annual prices are worth one line of arithmetic too. Twelve months of Standard at the early-bird rate is $216; the published yearly price is $195, about 9.7% less (our arithmetic). Against the $25 list price, the yearly figure is 35% less. Which of those two comparisons is the real one depends entirely on whether the early-bird rate survives, and the end date of the early-bird pricing is officially not published. We would not build a twelve-month forecast on a promotional monthly rate whose expiry the vendor has not stated — we watched a different Chinese vendor quietly end a promotion this same month, and the only signal was a banner disappearing from a documentation page nobody was watching.

The $18 collision, and why the two $18 plans cannot actually be compared

Here is the finding that made this worth writing. Two subscriptions, same price point, same five-hour-plus-rolling-window structure, both reachable from Claude Code. You would expect the comparison to be straightforward. It is not possible at all, and the reason is a disclosure asymmetry rather than a product difference.

The GLM Coding Plan publishes a Token Allowance table — the unit is tokens, which is the unit your code is measured in. Alibaba’s Token Plan publishes credits, and the per-model credit multipliers are officially not published in the announcement we read. Without a multiplier you cannot convert 10,000 weekly credits into any number of tokens, for any model, at any confidence. So the honest answer to “which $18 plan gives you more?” is: one of them tells you what you are buying and the other does not, and that asymmetry is the most decision-relevant fact available.

We want to be precise about what we are criticising. This is not a claim that the Alibaba plan is worse value — it may well be better, and we have no way to know. It is a claim that a unit you cannot convert is a unit you cannot budget in, and that a buyer comparing these two on price alone is comparing two numbers that happen to be equal and mean different things. If you are running the cross-vendor comparison for real, our AI coding assistant pricing breakdown lays out where the published units diverge across the whole market, and it is a wider problem than these two vendors.

Off-peak is a second lever, it is published as a multiplier, and it stacks

Between 22:00 and 08:00 Beijing time, the vendor publishes reduced credit consumption: 0.2x for Qwen3.7-Max and 0.4x for Qwen3.7-Plus, described in the announcement as stacking directly with Token Plan credits. Read that as a scheduling lever worth up to 5x on the same work for the first model and 2.5x for the second (our arithmetic on the published multipliers).

Two cautions. First, the off-peak multiplier is published for those two models and the multiplier for Qwen3.8-Max is officially not published — so if 3.8-Max is the model you actually want, the headline off-peak saving is not documented to apply to you. Second, a ten-hour window in China Standard Time is a real constraint on where you can live and still use it; it is the small hours in Europe and the working afternoon on the US west coast, which means the same subscription has materially different economics depending on your timezone. That is an eligibility condition wearing the costume of a discount, and price tables have no column for it.

The clauses that decide whether you are allowed to save the money

The announcement carries a set of conditions that will matter more than the price to some readers, and that appear nowhere in any comparison table we have seen:

• No refunds, and credits do not carry over between periods. Unused allowance is simply gone, which turns the weekly ceiling into a two-sided constraint: you are penalised for exceeding it and you are not credited for undershooting it.
• The plan is documented as not for automated scripts or backend service integrations. If your intended use is a CI job or a service calling the model on a schedule, this is not the product, whatever the arithmetic says.
• Access is through a dedicated key with its own prefix, distinct from a normal API key — the subscription and the pay-as-you-go account are separate doors, not two prices on the same door.
• The endpoint is Anthropic-compatible, and the announcement names Qwen Code, Claude Code, OpenClaw and OpenCode as compatible tools. Compatibility at the wire level is what makes this a genuine alternative rather than a different ecosystem.

The first and third of those are the ones we would put in front of a team lead. A subscription you cannot point your automation at, bought with a key that does not work where your existing key works, is not a drop-in substitution for API spend no matter how the per-credit arithmetic comes out. This is the same pattern we have documented repeatedly on this site: the rate card is one term in the bill, and the other terms are eligibility and usage pattern, and only the first one gets tabulated.

One thing we are deliberately not saying

The announcement describes Qwen3.8-Max-Preview as available exclusively on the Token Plan. We are not repeating that as a current fact, because the same vendor’s API pricing page lists qwen3.8-max as generally available, with a last-updated stamp of 11 September 2026. A July announcement and a September pricing page disagree, and when vendor sources conflict we report the conflict rather than picking the one that makes the better sentence. Do not plan around “3.8-Max is subscription-only.”

What to actually do with this

If you are choosing between the GLM Coding Plan, this Alibaba Token Plan, and staying on Claude Code, the arithmetic above does not decide it for you, and pretending otherwise would be the mistake this page exists to avoid. What decides it is three measurements you can take this week at no cost:

1. Your weekly, not your peak. Every plan in this class advertises a short window and is limited by a long one, at roughly ten to one. Pull a week of your own usage and compare it against the weekly ceiling, then ignore the five-hour figure entirely.
2. Whether your workload is even eligible. Automated scripts and backend integrations are excluded here. If more than a small fraction of your calls are unattended, the subscription comparison is moot and you are a pay-as-you-go customer whatever the price says.
3. Whether you can actually run in the discount window. A 0.2x multiplier between 22:00 and 08:00 Beijing time is either the largest lever on this page or entirely unavailable to you, and which one depends on your timezone rather than your architecture.

And one structural point worth carrying away: when a subscription is priced in a vendor-defined unit with undisclosed conversion rates, the per-unit arithmetic is theatre. You can compute $0.00180 per credit to five decimal places and still not know what a credit buys. If you want an approach that does not depend on the vendor’s unit at all, the model-downgrade route covered in our Claude Code alternatives guide is measured in the only unit that survives translation: total cost per completed task. And if what is actually driving your bill is hitting a ceiling rather than a rate, our breakdown of how usage limits reset is the more useful page — ceilings and rate cards fail in different ways and want different fixes.

Qualifiers, in order of how much damage they do

• Every figure above is the announcement blog’s, not the checkout page’s. We did not fetch the purchase flow. Where a blog and a checkout page disagree, the checkout page is what you will be charged.
• The early-bird monthly prices are promotional and their end date is officially not published. Every per-credit number in this section is computed from a promotional price and moves the moment that promotion ends.
• Credits are not tokens, and the conversion is not published. This invalidates any attempt to compare this plan against a per-million-token rate card, including ours. We have not attempted one and you should distrust any that you find.
• Cross-vendor per-token comparison is structurally unsound anyway. Anthropic has stated that Claude 4.7 and later use a different tokenizer that produces roughly 30% more tokens for the same text, so identical prices per million tokens are not identical prices for identical work. Read every cross-vendor multiple on this site as an order of magnitude, never a precise ratio.
• Concurrent-agent counts are published as ranges (1 to 2, 3 to 4, 6 to 8) and the vendor does not publish what determines where in the range you land.

Update, 23 September 2026: a third vendor now publishes an official Claude Code setup, and its default model mapping is the expensive one

This page has been tracking a specific claim: that you can point Claude Code at somebody other than Anthropic and keep working. As of 23 September 2026 there are three vendors shipping both an Anthropic-compatible endpoint and their own documented Claude Code configuration — Z.AI, Alibaba, and now DeepSeek, whose integration page we read for the first time today at api-docs.deepseek.com/quick_start/agent_integrations/claude_code.

The endpoint is https://api.deepseek.com/anthropic. That single fact is worth pausing on before the pricing: three separate vendors have now decided that the cheapest route to a coding-agent user is to implement somebody else’s API surface rather than persuade anyone to adopt theirs. Whatever else is true about this market, the client has stopped being a differentiator and become an integration target.

The mapping is a pricing decision wearing a compatibility label

Three lines from the integration page, quoted exactly:

Models starting with claude-opus are mapped to deepseek-v4-pro

Models starting with claude-haiku or claude-sonnet are mapped to deepseek-flash

The claude-opus mapping points to deepseek-v4-pro, which is billed at the V4 Pro price.

Read that as a bill rather than a routing table. Against the published price list, deepseek-v4-pro costs $0.66 off-peak and $1.32 peak for cache-miss input, where deepseek-flash costs $0.15 and $0.30 — 4.4x either way (our arithmetic). Output is 3.3x, and the cache-hit row is 7.3x ($0.022 against $0.003 off-peak, our arithmetic).

So the question of which price column you land in is decided by whether the string in your session begins with “claude-opus”. Not by what you asked for, not by task difficulty, not by anything you would think of as a purchasing decision.

And the vendor’s own quickstart overrides the vendor’s own default

Here is the part that makes this worth a section rather than a footnote. The environment block DeepSeek publishes on that same page sets:

• ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN
• ANTHROPIC_MODEL = deepseek-flash[1m]
• ANTHROPIC_DEFAULT_OPUS_MODEL = deepseek-flash[1m]
• ANTHROPIC_DEFAULT_SONNET_MODEL = deepseek-flash[1m]
• ANTHROPIC_DEFAULT_HAIKU_MODEL = deepseek-flash
• CLAUDE_CODE_SUBAGENT_MODEL = deepseek-flash
• CLAUDE_CODE_EFFORT_LEVEL = max
• CLAUDE_CODE_AUTO_COMPACT_WINDOW = 786432

The third line points the Opus slot at Flash. The documented happy path deliberately steers away from the documented default mapping, and the default mapping is the 4.4x side of that disagreement.

We are not calling this a trap, and we do not think it is one: the override is printed in the vendor’s own quickstart, and the mapping rule is stated in plain English a few paragraphs away. But the failure mode is real and specific. Set only the base URL and the auth token — which is what “point Claude Code at another endpoint” means to most people, and what most third-party write-ups show — and every Opus-class request in your session lands on the more expensive model, at 4.4x input, with nothing in the terminal indicating that it happened. You find out on the invoice, in a currency and a billing period that make attribution hard.

The cheapest thing anyone reading this page can do in the next ten minutes is open their own environment block and check which of the two paths they are on. If ANTHROPIC_DEFAULT_OPUS_MODEL is unset, you are on the expensive one.

Two of the other variables repay a second look. CLAUDE_CODE_AUTO_COMPACT_WINDOW at 786,432 is exactly 75% of the 1,048,576-token context window (our arithmetic) — compaction configured to fire at three quarters of the window rather than at its edge, which is a meaningfully more conservative setting than most defaults and has a direct cost consequence, since compaction events rewrite context and rewritten context is cache-miss input. CLAUDE_CODE_EFFORT_LEVEL at max pushes the other way. The vendor is not hiding either; it is simply that the two together encode an opinion about the cost/quality trade-off that you inherit silently by copying the block.

One more line from the page: Web Search works natively inside Claude Code against this endpoint, but “additional model token costs will be incurred”. That is the double charge we describe on our AI coding assistant pricing page — a fetched result is billed once on retrieval and again on every turn it remains in context.

Three subscriptions and one meter, and the gap is structural rather than arithmetic

The GLM Coding Plan and Alibaba’s Model Studio Token Plan both sell buckets — a five-hour allowance and a weekly allowance — from around $18 a month. DeepSeek sells no subscription at all. It is metered per token, with the time-of-day multiplier described on our pricing page, and nothing else.

There is no arithmetic that converts “$18 a month of buckets” into “$0.15 per million off-peak input tokens” without knowing your own volume. This is the same conclusion we reached on 22 September comparing the two $18 plans to each other, arrived at from the opposite direction: there, the blocker was that one vendor publishes token allowances and the other publishes credits with no conversion. Here, both sides publish honestly and in full, and they are still not comparable, because one is pricing a ceiling and the other is pricing a rate.

The single measurement that makes all three quotable is your weekly token volume, split into cache hits and cache misses. Until you have it, every comparison of these products — including every one published anywhere this month — is a comparison of two numbers that answer different questions.

The difference that survives the unit mismatch

Even without the conversion, one thing does compare cleanly, and it is not price. It is failure mode.

A subscription’s worst case is a wall. A metered API’s worst case is a bill. Buckets fail closed: you hit the weekly ceiling, the agent stops, you are annoyed, the cost is bounded and known in advance. Meters fail open: a runaway loop at 03:00 does not stop, and the first signal is financial and arrives later. Neither is better in general. But they are different products in a way that no price comparison surfaces, and which one you want depends on whether an interrupted Friday afternoon or an unbounded Saturday invoice is the thing you can least afford.

The practical consequence for anyone actually running a metered endpoint: the spend cap is not an optional safety feature, it is the component you are buying to replace the wall the subscription would have given you for free. Set it before the first real session, not after the first surprising one.

Off-peak, and a pattern that now spans two vendors

DeepSeek’s peak window is 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays, with everything else at half price. Converted, that puts peak squarely on the Chinese working day and entirely outside the American one (our arithmetic). Alibaba’s discount window — 22:00–08:00 Beijing, which we wrote up here yesterday — points the same way from the other side.

Two vendors, two opposite mechanisms, one outcome: the cheap hours are the hours their domestic users are asleep. For a reader outside China evaluating either product, that is a quiet tailwind worth an explicit line in the business case; for a reader inside China, it is the reverse, and the same headline rate means something close to twice as much money. Full working through on our pricing page.

Benchmarks: vendor claim, not used in any comparison here

DeepSeek’s changelog publishes the following for V4.1-Flash: GPQA Diamond 90.9, HLE 36.8, Codeforces 3471, Terminal-Bench 2.1 90.6, DeepSWE v1.1 74.2, CyberGym 88.1. Every one of those is a vendor claim taken from the vendor’s own changelog. We have seen no third-party reproduction, and we do not use vendor benchmark figures in any comparison, recommendation or arithmetic on this site. They are reproduced here only so that you know what the vendor is asserting when you go and check it.

What to do this week

1. Check ANTHROPIC_DEFAULT_OPUS_MODEL in whatever environment you already have. Unset means the expensive mapping. This is a one-line check with a 4.4x consequence on one row of your bill.
2. Measure your weekly token volume once, split by cache hit and miss. That number is what makes the subscription-versus-meter question answerable at all; without it you are comparing a ceiling to a rate.
3. Decide which failure mode you want before you decide which price you want. If the honest answer is “I cannot have a surprise invoice”, a bucket plan wins even at a worse headline rate, and vice versa.
4. Set a spend cap on day one if you go metered, and treat it as load-bearing rather than optional.
5. Check whether your working hours fall in the peak window before quoting yourself any DeepSeek figure. Between peak and off-peak the same table says two different things, and the difference is 2x.

Related reading: for what moved on the Anthropic and OpenAI side in the same week — Claude Opus 5.5, GPT-6 Sol and GPT-6 Luna — see our Codex vs Claude Code comparison. For the wider field of Anthropic-compatible endpoints and what each one actually supports, see Claude Code alternatives.

Update, 24 September 2026: a fourth vendor now sells a Claude Code subscription, and it is the first one priced in agents rather than tokens

When this page was first written, the question it answered was whether you could run Claude Code against anything other than Anthropic. That question is closed. As of our reading on 24 September 2026 there are four vendors with an official, documented path, and the newest of them prices its plan in a unit none of the others use.

MiniMax now publishes a subscription it calls the Token Plan. Its own documentation describes it as a successor to a product we had been tracking under a different name:

extends upon our former Coding Plan

That rename is not cosmetic, and we come back to it below. First the numbers, read from the vendor’s own pricing page on 24 September 2026.

The tiers

Tier	Price per month	Agents
Plus	$22	3–4
Max	$55	4–5
Ultra	$132	6–7

The plan covers M3 and M2.7 alongside image and speech models, and the vendor’s integration guides name Claude Code, Cursor, OpenClaw, TRAE and Hermes Agent. Windows are described as 5-hour rolling and weekly windows. Credits, sold separately, are priced at 1,000 credits to the dollar with a 365-day life.

The ladder charges you more per agent the higher you climb

Every volume price on this page works the same way: you buy more, the unit gets cheaper. That is what a tier ladder is for. MiniMax’s ladder does the opposite, and it does it consistently at both ends of every band.

Take the top of each band first. Plus is $22 for up to 4 agents, which is $5.50 per agent. Max is $55 for up to 5, which is $11.00. Ultra is $132 for up to 7, which is $18.86 (our arithmetic). Now take the bottom of each band, which is the number you are guaranteed rather than the number you might get: $22 for 3 is $7.33, $55 for 4 is $13.75, $132 for 6 is $22.00 (our arithmetic).

Read either column and the per-agent price roughly triples from the cheapest tier to the most expensive one. Moving from Plus to Max costs 2.5 times the money for one more agent at the band edge. Moving from Max to Ultra costs 2.4 times the money for two more (our arithmetic).

We want to be careful about what this does and does not mean, because it would be easy to write it up as a scandal and it is not one. The vendor does not define “agent” as a billing unit anywhere we could find, so our per-agent figures are arithmetic performed on a unit the vendor itself does not price. The tiers almost certainly differ in ways the agent count does not capture — window size being the obvious candidate, since the allowance inside each 5-hour and weekly window is officially not published. The honest statement is narrower and still useful: the one number MiniMax does publish for each tier gets worse per dollar as you go up, so if that number is the one you are shopping on, the cheapest tier is the best value and the ladder gives you no reason to climb it.

Which raises the question of what the ladder is actually selling, and the answer is almost certainly the thing that is not published. That is worth saying plainly: on this plan the advertised unit is the one that gets worse as you pay more, and the unit that presumably gets better is the one you cannot see.

An agent is a concurrency unit, not a consumption unit

This is the part that matters for anyone comparing the Token Plan against the other three vendors on this page. A token allowance tells you how much work you may do. An agent count tells you how many things may run at once. They are not the same quantity and no arithmetic converts between them, because one is a rate ceiling and the other is a total.

Three agents running an hour a day and three agents running around the clock are the same number on this price list and roughly a twentyfold difference in work done. Conversely, a single agent in a tight loop can consume more than six agents doing careful review. If you are choosing a tier by counting the seats on your team, you are shopping on the wrong axis; the agent count constrains fan-out, and fan-out is a property of your architecture, not your headcount.

We flagged the same distinction from the other direction when DeepSeek published concurrency as a model attribute while subscription vendors publish it as a plan attribute. MiniMax has now gone one step further and made concurrency the plan attribute — the headline unit, not a footnote to it.

Four vendors, four denominators

With MiniMax added, the four documented ways to point Claude Code at something other than Anthropic now bill in four different units:

• Z.AI (GLM) — prompts inside a rolling window.
• Alibaba (Qwen) — a token allowance.
• DeepSeek — metered per token, with a time-of-day multiplier.
• MiniMax — agents, plus separately-sold credits at 1,000 to the dollar.

We have written before that the denominator is one of the columns that decides your bill without appearing on the price list. Four vendors into this market, it is no longer a hidden column — it is the product differentiator. Not one of these four can be compared to another without you supplying your own usage data, and that is not an oversight that will be fixed in a later revision. A vendor that publishes a unit you can convert has handed you a shopping comparison against its competitors. None of them has done it, and the newest entrant has chosen the least convertible unit yet.

MiniMax is the only one of the four where you can compute a break-even — and it still will not tell you what you want to know

There is one genuinely useful thing about MiniMax’s disclosure that none of the other three offer: the same vendor publishes both a subscription and a pay-as-you-go rate card for the same models. M3 on the meter is $0.30 per million input tokens and $1.20 output at or below 512k context, with cached reads at $0.06; above 512k those become $0.60 and $2.40. Priority service is 1.5x.

So you can ask what $22 buys on the meter. Input-only, $22 at $0.30 per million is roughly 73 million tokens a month, about 2.4 million a day. On a one-to-one input-output mix at a blended $1.50 per million, it is roughly 15 million tokens a month (our arithmetic). Those are real numbers and you can check them against your own logs today.

But notice what that calculation actually established. It tells you what the money buys on the meter. It does not tell you what the money buys on the plan, because the plan’s allowance is officially not published. So the comparison you can perform is “meter versus meter”, and the comparison you need — “plan versus meter” — is the one the disclosure does not support.

This is worth stating as a general rule, because it will recur: when a vendor publishes a subscription price and withholds the subscription allowance, the only comparison it has made possible is one the subscription wins by construction. If the plan gave you less than $22 of metered usage, publishing the allowance would show that immediately. The absence of the number is not evidence that the plan is bad value. It is evidence that you cannot find out before you buy, which is a different problem and the one you should price in.

The usable version, then, is a floor rather than a verdict: $22 a month is worth taking seriously if and only if you would otherwise spend more than about $22 a month at list rate, and you can compute that side from your existing bill without MiniMax’s help. If your current metered spend on comparable work is $4, no subscription on this page is your answer regardless of what its allowance turns out to be.

The rename points the other way from Anthropic and OpenAI

A product called the Coding Plan became a product called the Token Plan, and its coverage now spans text, image and speech models rather than coding models. Meanwhile Anthropic and OpenAI have moved in the opposite direction, giving their coding clients their own dedicated limit structures separate from general API usage.

We are not going to tell you what that means about anyone’s strategy, because neither page gives a reason and we do not publish inferences about intent as though they were vendor statements. The practical consequence is concrete enough without a motive: a plan that is no longer coding-specific is a plan whose limits are shared with whatever else you point at it. If the same subscription is backing an image pipeline, that pipeline and your coding agents are drawing on the same 5-hour window. On the coding-specific plans they would not be.

The five-hour rolling window is now an industry convention that nobody standardised

MiniMax describes 5-hour rolling and weekly windows. That is the same two-level shape Anthropic uses and the same 5-hour period Z.AI uses. Three vendors, one window length, no standards body and no stated reason anywhere.

For a buyer this convergence is mildly good news and a significant trap. Good news, because the shape of the constraint is now portable: if you have learned how a 5-hour rolling window behaves on one vendor — that it is rolling rather than resetting, so a burst does not clear at a fixed hour — that knowledge carries. The trap is that the shape being identical makes it feel like the size is comparable, and the size is exactly what MiniMax does not publish. Two vendors can both sell you “a 5-hour window” and differ by an order of magnitude in what fits inside it.

A note on Anthropic’s own price page, because it affects what we said earlier on this page

Separately from MiniMax, one thing changed on Anthropic’s pricing documentation between our reading on 23 September and our reading on 24 September 2026. The prices did not change. Sonnet 5 remains $2 and $10 per million with cached input at $0.20; Opus 5.5 remains $4 and $20 with cached input at $0.20 and the 0.05x footnote; the 1.1x regional multiplier, the long-context provision and the tokenizer note are all still there.

What is gone is a sentence. The page previously carried an explicit statement that a previously scheduled increase to $3 and $15 would not occur. As of 24 September 2026 that sentence is no longer on the page, and the rate table has been restructured into four highlighted rows plus an “Additional models” section.

We are not going to tell you that this means a price rise is coming, and you should be suspicious of anyone who does. A sentence disappearing from a documentation page is consistent with an impending change and equally consistent with a routine cleanup of a note that had done its job. We do not know which, the page does not say, and guessing would be exactly the kind of number-shaped speculation this site exists to avoid. The verified facts are: the sentence was there on 23 September, it is not there on 24 September, and every price is unchanged.

The reason we are flagging it here rather than burying it is that this page quoted that sentence, and so did five others on this site. Those quotations were accurate when written and are no longer verifiable against the source. We are leaving them in place with this correction attached rather than editing them away, for the same reason we corrected ourselves in public when a source we had written off turned out to be readable: a site that quietly deletes its own outdated claims is asking you to trust a record it has been editing.

There is a reusable lesson in it, and it is the most portable thing in this update. A pricing page is two different kinds of document at once. It is a list of numbers, which the vendor maintains because customers check them. And it is a set of promises in prose, which nobody diffs. Numbers get changelogs. Sentences get deleted. If a procurement case of yours rests on something a vendor said rather than something a vendor charges, archive the sentence with the date you read it, because it is the half of the page that can vanish without leaving a mark.

What to actually do this week

1. Work out your own metered spend before you look at any of the four plans. It is the only number in this comparison that is fully knowable, it is already in your billing history, and every subscription decision on this page reduces to comparing it against a monthly fee.
2. If you are considering MiniMax, shop the bottom tier. The published unit gets worse per dollar as you climb, so unless you have a concrete fan-out requirement above three concurrent agents, the ladder is giving you no stated reason to go up it.
3. Count your agents by architecture, not by team size. The tier constrains how many things run at once. If your setup fans out to six workers for one developer, you are an Ultra customer with a headcount of one.
4. If a subscription is going to back more than coding, check whose window it shares. The Token Plan covers image and speech models on the same plan; a general-purpose subscription behind a coding agent is a shared ceiling, and you find out it was shared at the worst moment.
5. Go and archive the vendor sentences your budget depends on. Take ten minutes, screenshot or save the paragraphs — not the tables, the paragraphs — and date them. We have just watched one disappear.

For the full cross-vendor rate tables, including DeepSeek’s time-of-day multiplier and how it lands differently depending on which timezone your team works in, see our AI coding assistant pricing comparison. For how Anthropic’s own cached-input pricing stopped being derivable from its input pricing — the change that makes any spreadsheet built on a flat 0.1x cache multiplier silently wrong — see our Codex vs Claude Code comparison. And for the wider set of clients and back ends, not just the four with official Anthropic-compatible endpoints, see our roundup of Claude Code alternatives.

All prices in this section are vendor list prices read on 24 September 2026. Figures identified as our arithmetic are our own calculations on those published numbers, not vendor claims. Where a vendor does not publish a figure we say so rather than estimating it.