Last verified: 30 July 2026. Every number on this page is quoted from the vendor’s own documentation, with the URL. Where a vendor does not publish a number, this page says Not published and lists what was checked instead of guessing. Where two official pages contradict each other, both are shown and neither is declared correct.
The short answer
“5x” and “20x” are not marketing shorthand invented by reviewers. Both Anthropic and OpenAI use those exact multipliers in their own documentation, and both use them for their top coding tiers. That is where the similarity stops.
Anthropic’s multiplier is defined per session. OpenAI’s is defined against a published per-model range. Anthropic publishes no absolute message counts at any tier; OpenAI publishes them for Codex in fine per-model detail while publishing almost nothing for chat. So the two “20x” figures are not the same kind of claim, and no arithmetic converts one into the other.
If you only read one section, read why a cross-vendor limits table cannot be built. It is the reason most comparison articles on this topic contain numbers that no vendor ever published.
What Anthropic publishes
Anthropic’s Max plan page defines both multipliers in one sentence each. Verbatim:
“Max 5x provides 5 times more usage per session than the Pro plan”
“Max 20x provides 20 times more usage per session than the Pro plan.”
Source: support.claude.com — What is the Max plan?
Read those sentences closely. The unit being multiplied is “usage per session”, and the baseline is the Pro plan. Anthropic does not state what one unit of “usage” is, and does not state how many of them Pro gets. So the multiplier is exact and the thing it multiplies is undefined.
Because the Pro plan is the baseline that both multipliers refer to, it is worth knowing exactly what Anthropic does and does not publish about Pro itself. We went through ten official pages on that question in Claude Pro usage limits.
The session window itself is documented. Anthropic’s Claude Code cost documentation describes the allowance as one that “resets on a rolling five-hour window and a weekly window”. That gives you the shape: a short rolling window that recovers several times a day, plus a longer ceiling that does not.
What about an absolute number of messages? Not published. Checked: the Max plan article above, usage limit best practices (last updated 2 June 2026), and code.claude.com/docs/en/costs. The best-practices article says only that the number of messages varies by plan, and points readers elsewhere for specifics; the pages it points to do not carry a count either. Any article that tells you Max 5x is “about 225 messages per five hours” is quoting something Anthropic has not published.
What “5x” and “20x” multiply — and what they don’t
This is the single most consequential detail on this page, and it follows directly from the wording above.
Anthropic’s Max page states that Max plans carry two weekly limits in addition to the session window:
“Max plans also have two weekly usage limits: one that applies across all models and another for Sonnet models only.”
The multiplier sentences attach to “usage per session”. They do not attach to either weekly limit. Whether 20x also scales the weekly ceilings, or whether the weekly ceilings are set independently of tier, is Not published on any of the three pages checked above.
The practical consequence: if your bottleneck is the five-hour window — you hit a wall mid-afternoon and wait — the multiplier is the number that matters and buying up should help proportionally. If your bottleneck is the weekly ceiling — you run fine Monday through Wednesday and stall Thursday — the documentation does not promise that the multiplier helps you at all. Those are different purchases, and the marketing number only describes the first one.
Anthropic’s own docs disagree on which models the second weekly limit covers
The Max plan page, quoted above, says the second weekly limit applies to “Sonnet models only”. But Claude Code’s own error documentation describes a model-specific limit message reading “You’ve hit your Opus limit”, and the cost documentation discusses that same message as the model-specific ceiling a developer runs into, noting that switching models with /model “does keep the developer working after the model-specific ‘You’ve hit your Opus limit’ message”.
One official page names Sonnet. Another official page and the product’s own error string name Opus. Both are current Anthropic documentation as of 30 July 2026. This page does not adjudicate which is right — it is flagged here because if you are choosing a tier based on which model you actually use, this is the sentence you would be relying on, and it is not stable across Anthropic’s own surfaces.
One pool, several apps
A detail that changes the maths for anyone using both the chat app and the CLI: on Anthropic’s Team and Enterprise plans the documentation is explicit that the allowance “is shared with Claude chat and Cowork”, drawing from a single per-seat allowance on the same five-hour and weekly windows. Heavy chat use and heavy Claude Code use are not separate budgets there.
The equivalent statement for individual Pro and Max plans is Not published in the pages checked. Do not assume either way from the Team wording.
What OpenAI publishes: the chat / Codex split
OpenAI is close to the mirror image of Anthropic. Its consumer chat limits are largely unpublished, while its coding-tool limits are published in more granular detail than any other vendor on this page.
The multiplier terminology is official. OpenAI’s help centre states:
“Pro $100 unlocks 5x higher usage than Plus, while Pro $200 unlocks 20x”
Source: help.openai.com — About ChatGPT Pro tiers.
And it is not just prose. On developers.openai.com/codex/pricing, the usage-limit table’s column headers are literally Plus, Pro 5x, Pro 20x, Business, API Key. The consumer-facing chatgpt.com/codex/pricing phrases it as “Get 5x or 20x more usage than Plus”.
If you have seen a claim that OpenAI now advertises 10x rather than 5x, here is where that comes from. The developers.openai.com page carries a promotional line: “Double your normal Codex usage on the $100/month tier until May 31, 2026.” During that promotion the $100 tier temporarily behaved as 10x. That date has passed — it expired two months before this page was written — yet the banner is still displayed on the live page. The current, unpromoted multipliers on all three OpenAI surfaces checked are 5x and 20x.
The Codex per-model limits, as published
This is the one place where a vendor publishes hard numbers per tier. From developers.openai.com/codex/pricing, “Local Messages” per 5 hours:
| Model | Plus | Pro 5x | Pro 20x | Business |
|---|---|---|---|---|
| GPT-5.5 | 15–80 | 80–400 | 300–1600 | 15–80 |
| GPT-5.4 | 20–100 | 100–500 | 400–2000 | 20–100 |
| GPT-5.4-mini | 60–350 | 300–1750 | 1200–7000 | 60–350 |
| GPT-5.3-Codex | 30–150 | 150–750 | 600–3000 | 30–150 |
Two things are worth noticing before you use this table.
First, every cell is a range, and the ranges are wide — Plus on GPT-5.5 is anywhere from 15 to 80 messages in five hours, a factor of five between the floor and the ceiling. Whatever determines where you land inside that range is not published. So even OpenAI’s published numbers do not let you predict your own ceiling; they bound it.
Second, the multiplier and the table do not perfectly agree. On GPT-5.5, Pro 5x is 80–400 against Plus at 15–80: the ceiling is exactly 5x, the floor is 5.3x. On GPT-5.4, Pro 20x is 400–2000 against Plus at 20–100 — exactly 20x at both ends. On GPT-5.5, Pro 20x is 300–1600 against 15–80, which is 20x at both ends. The arithmetic mostly holds, and where it does not, the discrepancy is small. That is worth stating plainly because it is the only vendor on this page where you can check the marketing multiplier against a published number at all.
OpenAI’s two Codex pricing pages list different models
As of 30 July 2026, developers.openai.com/codex/pricing tops out at GPT-5.5 and shows the tier-column table above. chatgpt.com/codex/pricing shows a different model lineup — GPT-5.6 Sol at 15–90, GPT-5.6 Terra at 20–110, and GPT-5.6 Luna at 50–280 local messages per 5 hours — with columns for Local Messages, Cloud Tasks and Code Reviews rather than for tiers, and does not break the newer models out per tier.
The developer page also still displays the expired May promotion. Taken together, that page appears to be the staler of the two, but OpenAI does not label either as superseded, so both are quoted here as found. If you are checking your own limit for a GPT-5.6 model, the tier multipliers are the only official guidance — the per-tier numbers for those models are Not published.
What a Codex message actually spends: credits, not messages
OpenAI’s help page for using Codex on a ChatGPT plan publishes no numbers at all. It says only that “Codex is included across ChatGPT plans, including Free and Go” and that “Usage limits vary by plan,” then points at the developer pricing page. The reason the message counts above are ranges rather than fixed numbers sits one layer down: Codex does not meter messages. It meters tokens, converts them to credits, and derives the message ranges from an assumed message size. OpenAI states the dependency plainly — the amount a task uses “varies with the size and complexity of the work, the model, and where the task runs.”
The credit rates OpenAI publishes on the Codex pricing page, per 1M tokens:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-5.5 | 125 | 12.50 | 750 |
| GPT-5.4 | 62.50 | 6.250 | 375 |
| GPT-5.4-mini | 18.75 | 1.875 | 113 |
| GPT-5.3-Codex | 43.75 | 4.375 | 350 |
Credits per 1M tokens. Source: developers.openai.com/codex/pricing. The separate Codex rate card lists 12 models rather than 4, spanning 18.75 to 500 credits per 1M input tokens.
Two things follow from that table that a message count hides. Cached input costs roughly a tenth of fresh input on every model listed, so a long session that keeps re-reading the same files is cheaper than the same token volume split across fresh sessions. And output runs 6x input on GPT-5.5, so a task that writes a lot of code costs far more than one that reads a lot — even though both count as one message.
OpenAI’s own per-message estimates, for Plus and Pro: a local task is about 2 to 14 credits, a cloud task about 25, a 1024×1024 image 5 to 6, and a 1024×1536 image 7 to 8. The rate card’s legacy per-message averages run higher on the larger models — about 14 credits for a local task on GPT-5.6 Sol, about 34 for a cloud task on GPT-5.4.
The Codex weekly ceiling is not published
The pricing page carries one sentence that matters more than any figure on it: “The usage limits for local messages and cloud tasks share a five-hour window. Additional weekly limits may apply.”
That sentence does two jobs. It confirms local and cloud draw on one pool rather than two, so a heavy day of cloud tasks eats into the local message budget shown in the table above. And it confirms a weekly limit exists without saying how large it is.
We checked three OpenAI pages for that number: the Codex pricing page, the ChatGPT-plan help article, and the Codex rate card. None of them publishes a weekly figure. Not published. If you find a specific weekly Codex number quoted somewhere, check whether the vendor page it cites actually contains it. Neither Codex pricing page carries a visible last-updated date either, which is its own problem: you cannot tell how stale the ranges are, and the model names on them have changed more than once.
Anthropic is in the same position from the other direction. It names a weekly window and publishes no absolute figure for it. Both vendors publish the shape of the limit and withhold the size, which is why the comparison in the next section cannot be made numeric.
How to check what you have left, in each product
Neither vendor’s documentation puts this in one place, and it is more useful than any published range, because it reports your actual account rather than a band that spans a factor of five.
In Claude Code the command is /usage. It covers two windows — “Press d or w to switch between the last 24 hours and the last 7 days.” On Pro, Max, Team and Enterprise plans it goes further than a total: it “attributes recent usage to skills, subagents, plugins, and individual MCP servers,” and it “flags behaviors such as long context or cache misses” when one of them accounts for 10% or more of recent usage. That attribution is the most actionable number either product exposes, because it names which MCP server or subagent is quietly consuming the week.
One caveat Anthropic states directly: the figures are “approximate and computed from local session history on this machine,” so usage from another laptop or from claude.ai is not included in what you see. Two related commands: /cost shows running spend for the current session and is aimed at API-key users, and /usage-credits opens billing settings on Pro and Max, or organisation usage settings on Team and Enterprise.
Codex has no equivalent command. The corresponding surface is the Usage panel in Codex settings, where OpenAI says users “may also be able to view remaining credit, purchase credits, or manage auto-reload” — with access depending on workspace role. On a Business or Enterprise workspace, an ordinary member may not be able to see the balance at all.
So the two products answer different questions. Claude Code tells you where your usage went and not how much money is left. Codex tells you how much is left and not where it went. If you are trying to decide whether to upgrade, that asymmetry matters more than the published ranges: on one side you can fix the cause, on the other you can only watch the balance.
Claude Code weekly limits: what resets, and when
For Team and Enterprise, Anthropic describes the mechanism directly: each member’s usage “draws from a per-seat allowance that resets on a rolling five-hour window and a weekly window,” and “The allowance is shared with Claude chat and Cowork.”
Two consequences people miss. A long Claude chat session in the browser spends the same seat allowance as a coding session, so the limit you hit in the terminal may have been spent somewhere else. And because the window is rolling rather than calendar-aligned, there is no fixed hour at which it refills — capacity comes back five hours after the usage that consumed it, in pieces.
Claude Code rate limits are not the same thing as usage limits
The windows above are usage limits, attached to a subscription seat. If you run Claude Code against an API key, Bedrock or Vertex instead, the thing that stops you is a rate limit, expressed in tokens per minute and requests per minute and provisioned per organisation. Anthropic’s own planning guidance scales the per-user allocation down as headcount grows, because concurrency smooths out: roughly 200k to 300k TPM and 5 to 7 RPM per user for a team of one to five, falling to roughly 10k to 15k TPM and 0.25 to 0.35 RPM per user above 500. Hitting a rate limit produces a retry, not a lockout for the rest of the window, which is the practical difference.
Claude Code token limits: what actually burns tokens
Three figures Anthropic publishes move consumption more than model choice does. Prompt cache lifetime is “an hour on a subscription and drops to five minutes once you’re drawing on usage credits,” and is five minutes by default on an API key or cloud provider — so a subscription rewards leaving one session open where an API key rewards batching work into tight bursts. Agent teams “use approximately 7x more tokens than standard sessions when teammates run in plan mode,” which is the single largest multiplier in the docs. And background token usage is “typically under $0.04 per session,” so whatever is draining your week, it is not that.
For scale on the API side, Anthropic’s enterprise observation is “around $13 per developer per active day and $150-250 per developer per month,” with costs staying below $30 per active day for 90% of users. A Max subscription is below that monthly band, which is the actual arithmetic behind most upgrade decisions — not the message counts.
Additional sources checked for this section: help.openai.com/en/articles/11369540-using-codex-with-your-chatgpt-plan, help.openai.com/en/articles/20001106-codex-rate-card, developers.openai.com/codex/pricing, code.claude.com/docs/en/costs, support.claude.com/en/articles/14552983-models-usage-and-limits-in-claude-code. Checked 3 August 2026. No weekly Codex figure appears on any of them.
Why you cannot build a cross-vendor limits table
This is the finding that most comparison content on this topic gets wrong, and it is worth being blunt about: as of mid-2026 the five major AI coding vendors do not meter the same quantity, and three of them changed units within the last year. A table with a “messages per month” column for each vendor is not a comparison. It is a set of numbers that were converted using assumptions nobody published.
Here is what each vendor actually counts.
Anthropic counts an undefined internal unit of “usage” over a rolling five-hour window plus weekly windows, with tier expressed purely as a multiplier. No absolute figure at any tier.
OpenAI counts “local messages” per five hours for Codex, published as per-model ranges per tier — plus separate cloud-task and code-review counters. Chat-side limits are largely unpublished.
Google is the only vendor still using a clean daily request quota. Gemini Code Assist publishes “Maximum requests per user per day” of 1000 for the free individual tier, 1500 for Standard and for Google AI Pro, and 2000 for Enterprise and for Google AI Ultra. Critically, the documentation states these are “aggregated across all interactions with any model version or family” — so unlike OpenAI, a cheaper model does not buy you more requests. Source: developers.google.com — Gemini Code Assist quotas.
GitHub Copilot has moved to “premium requests”, where one action can consume more than one request: “The number of premium requests a feature consumes can vary depending on the feature and the AI model used.” Published allowances are 300 per month for Copilot Pro and 1500 per month for Copilot Pro+. The widely-circulated per-model multiplier table (0.33x for small models, 1x for the default, up to 10x) is now explicitly labelled legacy in GitHub’s own documentation, applying only to annual subscribers who remained on request-based billing after 1 June 2026. Source: docs.github.com — About premium requests.
Cursor has abandoned request counting entirely and now meters in dollars of inference at API prices: “$20 of API agent usage” on Pro, “$70 of API agent usage” on Pro+, and “$400 of API agent usage” on Ultra, each plus bonus usage. Source: cursor.com/docs/account/pricing.
Line those up and the problem is obvious. Anthropic’s unit is undefined, OpenAI’s is a bounded range, Google’s is a flat request, Copilot’s is a weighted request, and Cursor’s is a dollar. Converting between any two of them requires knowing the average token cost of your own work, which none of the vendors publish and which varies by an order of magnitude between a one-line autocomplete and a multi-file agent run.
The honest comparison is structural, not numeric. Which window will you hit — hourly, daily, weekly, monthly? Is your tier a multiplier or an absolute? Does a cheaper model buy you more headroom, or is the quota model-blind? Those questions have answers. “Which plan gives me more messages” does not.
How to tell which limit you actually hit
Because the vendors stack several windows, the message you see matters more than the number you were promised.
On Anthropic’s side, the documentation distinguishes three situations that look similar and mean different things. “You’ve hit your session limit” or “You’ve hit your weekly limit” is a plan usage window; the docs note these “windows are shared across all models, so switching models with /model doesn’t restore access”. A model-specific message such as “You’ve hit your Opus limit” is different — there, switching models does keep you working. And an auto-compact or context warning is not a usage limit at all; it means the conversation has grown near the model’s maximum input size.
That distinction is worth internalising before you upgrade, because only the first category is a capacity problem that money fixes. The second is solved by switching model, and the third by clearing context — neither is a reason to buy a higher tier.
What to check before you upgrade
Four questions, in the order that saves the most money.
Which window is actually stopping you? If you stall at the same time of day and recover a few hours later, that is the rolling window, and a tier multiplier addresses it. If you stall late in the week and do not recover until the week rolls over, that is a weekly ceiling, and no vendor on this page promises that its multiplier scales it.
Is your work concentrated in one model? On Google’s quota the model does not matter — the daily cap is aggregated across all of them. On OpenAI’s Codex table it matters a great deal: GPT-5.4-mini’s Plus allowance of 60–350 is roughly four times GPT-5.5’s 15–80. On Copilot the model is the multiplier. Moving one habitual task to a cheaper model can be worth more than a tier upgrade on three of the five vendors.
Are you spending your allowance on context rather than on work? Anthropic’s documentation is unusually direct on this: a long-running session re-sends the full conversation with every request, so “a one-line question in a session that has been open all day still draws usage for the whole conversation”. Clearing between unrelated tasks is free; a higher tier is not.
Is the number you are comparing one the vendor actually published? By the count on this page, the only absolute per-tier figures published anywhere are OpenAI’s Codex ranges, Google’s daily request caps, Copilot’s 300 and 1500 premium requests, and Cursor’s dollar allowances. Anthropic publishes none. Our Codex vs Claude Code cost breakdown works the same problem from the billing side. If a comparison shows you a tidy message count for Claude Max, that number was invented somewhere downstream of the vendor.
Sources, and what was checked
All pages below were fetched on 30 July 2026. Where a page shows a last-updated date, it is noted.
- Anthropic — What is the Max plan? (5x/20x per-session definitions; two weekly limits, “Sonnet models only”)
- Anthropic — Usage limit best practices (last updated 2 June 2026; no message counts)
- Anthropic — Manage costs effectively (five-hour and weekly windows; shared allowance on Team/Enterprise; “You’ve hit your Opus limit”)
- OpenAI — About ChatGPT Pro tiers (5x / 20x)
- OpenAI — Codex pricing (developer docs) (per-tier per-model table; expired May 2026 promotion still displayed)
- OpenAI — Codex pricing (consumer) (GPT-5.6 model lineup; “Get 5x or 20x more usage than Plus”)
- Google — Gemini Code Assist quotas (1000 / 1500 / 2000 requests per user per day, aggregated across models)
- GitHub — About premium requests (300 / 1500 per month; multiplier table marked legacy)
- Cursor — Pricing ($20 / $70 / $400 of API agent usage)
Two claims are deliberately absent from this page because no official source for them was found: a message count for any Claude tier, and a monthly session allowance for any Claude tier. Both circulate widely. Neither appears in the three Anthropic pages listed above.
Related on this site
- Claude Pro usage limits, and what Anthropic does not publish — the plan that every multiplier on this page is measured against.
- real monthly costs and where each one breaks — the same two tools compared on cost and failure points.
- AI coding assistant pricing in 2026 — what each tier costs per month across the same five vendors.
- Claude Code alternatives, with the break-even math — what to move to when the ceiling, not the price, is the problem.
- Best AI coding models compared — the model-level comparison behind these quotas.
If you have settled on Claude and only need to size your plan, see which Max tier to choose — 5x vs 20x.
For the timing side of all this, see when Claude usage limits reset: the five-hour session window, the account-assigned weekly reset, and what Anthropic still does not publish.
OpenAI has since retired the developer pricing URL and moved these tables. For the current per-tier figures, read Codex usage limits by tier, verified 21 August 2026.
Update, 5 September 2026: a seventh model, and the multipliers still hold
OpenAI added GPT-6 Astra to its published Codex tables at the start of September, above GPT-5.6 Sol. Since this article is built on the question of whether “5x” and “20x” mean anything precise, the new row is worth running through the same check.
Astra’s local-message allowance reads 5–45 on Plus, 25–225 on Pro 5x, and 100–900 on Pro 20x. Five times five is twenty-five and five times forty-five is two hundred and twenty-five; twenty times five is one hundred and twenty times forty-five is nine hundred. Both ends of both multipliers land exactly (our arithmetic). Across seven models and both ends of every range that is fourteen out of fourteen exact matches — OpenAI is still holding its tier arithmetic to the digit, which remains the sharpest contrast with Anthropic’s side, where no equivalent published table exists to check.
Two cautions before you read that as a reason to upgrade. First, Astra’s Plus floor is five messages per five hours — the tightest allowance anywhere on the card, half of Sol’s. A clean 20× multiple of a very small number is still a small number. Second, the multiplier applies to the message range, and OpenAI’s own footnote still says these “are not fixed message limits” and that weekly limits may also apply on top.
Separately, OpenAI cut GPT-5.6 Sol’s credit rate in the same window — input and cached input down 20%, output down a third — without changing any of the message ranges above. We work through what that implies in Codex usage limits, and what it does to the cross-vendor price gap in Claude Code vs Codex.
Read from OpenAI’s official Codex pricing documentation on 5 September 2026. Multiplier checks are our arithmetic on the published ranges.
Update, 14 September 2026: one side of this comparison publishes capacity numbers and the other does not
A temporary 50% increase to weekly Claude Code limits, announced on Anthropic’s developer channel on 19 August 2026 and extended twice, expired on 13 September 2026. We re-read Anthropic’s official support article on Claude and Claude Code usage limits on 14 September: it contains no numeric ceiling, no weekly figure, and no mention of the increase or its expiry. The change existed for roughly four weeks without ever entering the documentation.
That asymmetry is the thing to carry away from this page. OpenAI publishes per-five-hour message estimates and a credit rate card; those numbers have contradicted each other before, and we have documented it. Anthropic publishes no subscription capacity numbers at all. A published-but-inconsistent number can be audited; an unpublished one cannot. When you compare these two on capacity, you are comparing a table against an absence, and the correct entry for the absence is “unknown” rather than “comparable.”
The practical consequence is that for subscription capacity the vendor’s documentation is not the source of truth — your own usage meter is. We have applied the same correction on our Claude Pro usage limits page and our Codex usage limits page, both of which quoted the 19 August wording directly.
Update, 25 September 2026: one side of this comparison has a dated retirement 19 days out
This page names GPT-5.5 eight times and, until now, never carried the date that decides whether those mentions still describe anything. OpenAI’s pricing documentation, read this morning, states that GPT-5.5 retires from Codex on 14 October 2026. That is nineteen days from this update.
Nothing about the prices or multipliers above changes on that date. What changes is the denominator. Everything on this page compares two plans by what a fixed allowance buys, and a fixed allowance buys work only relative to a model’s rate. When the model is withdrawn and you are re-routed, the plan price, the allowance and the multipliers all stay exactly where they are while the amount of work you get moves.
How far it can move: the three credit tiers published this morning run 250 / 50 / 2.5 credits per million input tokens and 1,250 / 250 / 12.5 on output — a 100× span top to bottom on both, or 250× once the listed 2.5× “Fast” multiplier is applied (our arithmetic on the official rates). We are not computing your migration multiplier, because the model Codex routes to afterwards is officially not published. Anyone who gives you that number today has filled in the blank themselves.
A second date, and read the wording carefully because it is a floor and not an expiry: the page states that Sol’s promotional pricing is available “at least through November 21, 2026”. That promises no increase before that day and says nothing about the day after. Calendar it as a review, not as a migration.
Both dates are sentences on a pricing page rather than cells in its table — and sentences on pricing pages are the part that vanishes without a changelog. Copy them into your own notes today, with the date you read them. We walk through the credit-menu arithmetic and the cache-hit convention in full on Codex usage limits, and the head-to-head that already carried this date is Codex vs Claude Code.