On 1 June 2026 GitHub did something no other AI coding vendor has done: it started publishing, in dollars, exactly how much model usage each Copilot plan includes. Premium request units are gone. In their place is a credit worth precisely one cent, spent against published per-token API rates, with a stated monthly allowance for every tier from Pro to Enterprise.
Anthropic has not done this. Claude Code is included with Claude Pro, Max, Team and Enterprise, and the pricing page describes what you get in multiples — “5x or 20x more usage than Pro” — without ever stating what Pro itself is. Multiply an undisclosed number by twenty and you still have an undisclosed number.
So this comparison is lopsided in an unusual way. On one side there is an allowance you can put in a spreadsheet. On the other there is a product that many developers prefer, priced in a unit nobody outside Anthropic can see. This page puts every figure both companies actually publish into the same table, converts what can be converted, and marks plainly where the published record stops.
The short answer
If your question is “which one can I budget for,” the answer is GitHub Copilot, and it is not close. Copilot Pro is $10 a month and includes 1,500 AI credits, which is $15 of metered usage. Copilot Max is $100 a month and includes 20,000 credits, which is $200. Copilot Business is $19 per seat and includes 1,900 credits per seat, pooled across the organisation. Every one of those numbers comes from GitHub’s own billing documentation.
If your question is “which one is the better agent,” the published record cannot answer it, and anybody who tells you otherwise is quoting a benchmark that neither vendor ran. What we can tell you is that the two products are not the same shape. Copilot’s unlimited, unbilled tier is code completion inside your editor. Claude Code has no completion product at all — it is a terminal agent that reads your repository, runs commands and edits files. Those are different tools that happen to be sold to the same buyer.
And if your question is “can I have both,” the answer is partly yes, in a way most comparison pages miss: Anthropic’s Claude models are available inside Copilot. Sonnet, Opus and Haiku all appear in GitHub’s supported model list. What you cannot do is run the Claude Code CLI on a Copilot subscription. We come back to this below.
Every published price, side by side
These are list prices as printed by each vendor in August 2026. Copilot’s individual plans are billed monthly; Claude’s cheaper figures require annual prepayment.
| Plan | Price | Included usage, as published |
|---|---|---|
| Copilot Free | $0 | 2,000 code completions/month; an AI credit allowance (amount not published on the pages we checked); auto model selection only |
| Copilot Pro | $10/month | 1,500 AI credits ($15) |
| Copilot Pro+ | $39/month | 7,000 AI credits ($70) |
| Copilot Max | $100/month | 20,000 AI credits ($200) |
| Copilot Business | $19/seat/month | 1,900 credits/seat ($19), pooled org-wide |
| Copilot Enterprise | $39/seat/month | 3,900 credits/seat ($39), pooled org-wide |
| Claude Free | $0 | Claude Code not included |
| Claude Pro | $17/month annual ($200 upfront), $20 monthly | “More usage” than Free. Amount not published |
| Claude Max | From $100/month | “Choose 5x or 20x more usage than Pro.” Base amount not published; the 20x list price is not printed on the pricing page |
| Claude Team, standard seat | $20/seat/month annual, $25 monthly | “More usage than Pro.” Amount not published |
| Claude Team, premium seat | $100/seat/month annual, $125 monthly | “5x more usage than standard seats.” Amount not published |
| Claude Enterprise | $20/seat, annual only | “Usage cost scales with model and task” |
Read the right-hand column twice. Six rows contain a number. Five contain a multiplier or a sentence.
What a Copilot credit is, and what a dollar of it buys
GitHub’s billing documentation is unusually blunt: 1 AI credit = $0.01 USD. Credits are consumed by token count — input, output and cached tokens — priced at the published API rate for whichever model you selected. Pick a cheap model and your allowance stretches; leave Opus on by default and it does not.
Each individual plan’s allowance is split into two parts, and the distinction matters more than it first appears.
| Plan | Price | Base credits | Flex allotment | Total | Value | Ratio to price |
|---|---|---|---|---|---|---|
| Copilot Pro | $10 | 1,000 | 500 | 1,500 | $15 | 1.5x |
| Copilot Pro+ | $39 | 3,900 | 3,100 | 7,000 | $70 | 1.79x |
| Copilot Max | $100 | 10,000 | 10,000 | 20,000 | $200 | 2.0x |
Base credits, in GitHub’s words, “match with your subscription price and they never change.” The flex allotment is described as “a variable part of your included usage; it is designed to adapt as the economics of AI evolve.”
That sentence is the fine print of the whole scheme, and we have not seen another comparison page quote it. A third of Copilot Pro’s allowance and half of Max’s sits in a bucket the vendor has explicitly reserved the right to move. When GitHub first announced the transition, the stated mapping was simpler — Pro $10 buys $10 of credits, Pro+ $39 buys $39 — and the flex allotment is why the live figures are now higher. It can go the other way. Budget against the base; treat the flex as a discount that is currently being offered.
Business and Enterprise pool their credits
Organisation plans work differently in a way that flatters them. Credits accrue to the billing entity, not the individual, so an enterprise with 100 Copilot Business users gets a shared pool of 190,000 credits rather than 100 separate buckets of 1,900. Your two heaviest users do not starve while forty light users leave their allowance on the table.
Both organisation tiers ran a promotional allowance from 1 June to 1 September 2026: Business got 3,000 credits per seat instead of 1,900, Enterprise 7,000 instead of 3,900. That promotional period ended on 1 September 2026, and both tiers are now on the standard allowance — Business 1,900 credits per seat, Enterprise 3,900. At the promotional rate a $19 Business seat included $30 of usage, a 1.58x ratio; at the standard rate it is 1.0x. The section at the end of this piece works through what that changes for a rollout.
What Anthropic publishes instead
Anthropic does not publish an included allowance for any subscription tier. What it does publish is the mechanism and, in one specific document, some real observed spend.
The mechanism, from the pricing page: “Every plan has usage limits that reset on a rolling five-hour session window, and paid plans add weekly limits on top.” That is two ceilings, neither of them numbered. We took this apart in detail in our piece on what Claude Pro’s usage limits actually are, and the conclusion there was the same as here: the number does not exist in public.
The observed spend comes from Anthropic’s own cost-management documentation for Claude Code, and it is the most useful number either company has put in writing:
- About $13 per developer per active day. That is the average across enterprise deployments.
- $150–250 per developer per month. Same source, same population.
- Under $30 per active day for 90% of users. The tail above that is real but small.
- Agent teams use roughly 7x more tokens than standard sessions when teammates run in plan mode.
These are token costs at API rates, not subscription prices — they describe what the work consumes, not what you pay. But because Copilot credits are also metered on published API token rates, the two are close enough to compare, and the comparison is startling.
Putting both on the same axis
A Copilot Business seat includes $19 of metered usage per month. Anthropic’s published average for a Claude Code developer is around $13 of tokens per active day. If a Copilot Business developer worked at that same intensity, the included allowance would cover roughly a day and a half of active work per month. Against the $150–250 monthly figure, $19 is eight to thirteen percent of the bill.
Three caveats, and they matter. First, Copilot’s default routing does not put every request through a frontier model, and the cheap-model requests cost a fraction of the Opus-heavy sessions Anthropic’s average describes. Second, Copilot’s code completions — the feature most developers use most hours of the day — are not billed in credits at all and remain unlimited on every paid plan, so a large slice of a Copilot user’s actual usage never touches the allowance. Third, Anthropic’s figure is drawn from enterprise API deployments, which skew heavier than an individual on a $20 subscription.
What survives all three caveats is this: a published allowance in dollars lets you check the arithmetic and argue with it. A multiplier does not. If Anthropic published a per-tier token allowance tomorrow, this section would be two sentences long instead of six.
What actually competes with what
Price tables invite a category error. Copilot and Claude Code overlap, but they are not the same product, and the parts that overlap are not the parts that cost the most.
GitHub bills AI credits for Copilot Chat, Copilot CLI, the Copilot cloud agent, Copilot Spaces, Spark and third-party coding agents. It does not bill credits for code completions and next-edit suggestions, which stay unlimited on every paid plan. So Copilot’s flagship feature — the grey text that appears as you type — is free-at-the-margin, and the credit allowance is really the budget for its agentic half.
Claude Code has no completion product. It is a terminal agent: you describe a change, it reads the repository, runs commands, edits files and reports back. The nearest thing on the GitHub side is Copilot CLI and the cloud agent, both of which draw on credits. That is the honest head-to-head: Copilot CLI and the Copilot cloud agent versus Claude Code, with Copilot’s completions as a free extra on one side and nothing equivalent on the other.
If you want the same argument made against a different competitor, our Claude Code vs Cursor comparison covers the editor-native case, and our pricing overview puts every vendor’s list price in one place.
You can run Claude models inside Copilot
This is the fact that reframes the whole comparison, and it is buried in GitHub’s model reference rather than its marketing. Copilot’s supported model list includes Anthropic’s Claude Sonnet, Claude Opus and Claude Haiku alongside OpenAI, Google, xAI and Moonshot models. Paid Copilot plans can select among them; Copilot Free and Copilot Student get models through auto model selection only, and enterprise administrators can restrict the catalogue for their organisation.
So “Claude Code vs GitHub Copilot” is not “Anthropic’s models vs OpenAI’s models.” You can pay GitHub and still be talking to Claude — the tokens are simply metered against your Copilot credits at Anthropic’s published API rates rather than against a Claude subscription. What you cannot do is run the Claude Code CLI itself on a Copilot plan. Claude Code requires a Claude subscription or Console access; a Copilot seat does not unlock it. See our Claude Code alternatives roundup for the tools that do run on other billing.
What happens when you run out
The failure modes are where the two systems differ most, and where the published record is most complete on both sides.
| GitHub Copilot | Claude Code | |
|---|---|---|
| Reset cadence | Monthly, at 00:00:00 UTC on the first day of each calendar month | Rolling five-hour session window, plus a weekly window on paid plans |
| Rollover | None. “Unused credits are forfeited” | Not applicable — no stated allowance to carry |
| At the wall, individuals | Set a dollar budget and continue at published per-credit rates, upgrade, or wait for the reset | Wait for the window, or turn on usage credits and continue |
| At the wall, organisations | Additional usage is on by default: usage continues at published rates. Turn it off and usage is blocked until the pool refreshes | The seat allowance is the default ceiling; admins turn on usage credits and set spend limits per org, group or member |
| Overage price | $0.01 per credit, the same rate as included usage | Metered at API rates once drawing on usage credits |
| Model fallback | Removed. GitHub’s announcement states “Fallback experiences will no longer be available”; under the old premium-request scheme you could drop to a cheaper model and keep going | Switching models with /model does not restore a session or weekly window — those windows are shared across models |
| Hidden cost of overage | None documented | Prompt cache lifetime drops from one hour on a subscription to five minutes once you are drawing on usage credits |
That last row deserves a moment. On a Claude subscription the prompt cache lives for an hour, so stepping away for coffee and coming back costs you nothing. Once you cross into usage credits, the lifetime drops to five minutes, which means the first message after any real break reprocesses your entire context at full rate. The mechanism that lets you keep working past the limit is also the mechanism that makes each subsequent request more expensive. Neither vendor’s pricing page mentions this; it appears only in Anthropic’s cost documentation.
Sizing a team
For organisations, GitHub gives you a pooled credit figure and a budget control panel — user-level budgets, cost-centre budgets, organisation budgets and enterprise spending limits. Anthropic gives you something different but genuinely useful: per-user rate-limit recommendations by team size, for API and Console deployments.
| Team size | TPM per user | RPM per user |
|---|---|---|
| 1–5 users | 200k–300k | 5–7 |
| 5–20 users | 100k–150k | 2.5–3.5 |
| 20–50 users | 50k–75k | 1.25–1.75 |
| 50–100 users | 25k–35k | 0.62–0.87 |
| 100–500 users | 15k–20k | 0.37–0.47 |
| 500+ users | 10k–15k | 0.25–0.35 |
The per-user allocation falls as the team grows because fewer people use the tool concurrently in a large organisation. These limits apply at the organisation level, not per individual, so one developer can temporarily exceed their calculated share while others are idle. GitHub publishes no equivalent sizing table; its answer to the same question is the credit pool.
How to choose, concretely
If you need a number your finance team can approve, choose Copilot. Every tier states its included usage in dollars, overage is a flat cent per credit, and budgets can be capped at four levels. Nothing on the Claude side gives you this.
If completions are most of what you want, choose Copilot and stop reading. They are unlimited, unbilled on every paid plan, and Claude Code does not offer them.
If your work is agentic — long multi-file changes driven from a terminal — try Claude Code first and accept that you cannot forecast it. Anthropic’s own documentation says the average enterprise developer consumes $150–250 of tokens a month doing this, which no $10 or $19 seat covers on either platform. Start on Pro at $17–20, watch /usage, and move up only when you actually hit the wall.
If you are rolling out to a team, price the standard number, not the summer one. Copilot Business at 3,000 promotional credits per seat was a 1.58x ratio; at the standard 1,900 — which is what applies from 1 September 2026 onward — it is 1.0x. Anthropic’s Team standard seat is $20–25 with no stated allowance, and the premium seat is $100–125 for five times an unstated base.
If you want both, you can have most of both. A Copilot seat gives you completions plus Claude models in chat and CLI. A Claude subscription on top gives you Claude Code itself. That is roughly $10 plus $20 a month for an individual, and it is a defensible stack.
What we could not verify
In keeping with how we write every comparison on this site, here is what is missing rather than guessed.
- Claude’s numeric usage limits. No token, message or hour figure is published for Free, Pro, Max, Team or Enterprise. The multipliers are all that exist.
- The Claude Max 20x list price. The pricing page says “from $100/month” and offers a 5x or 20x choice without printing the second price.
- Copilot Free’s AI credit allowance. The documentation confirms Free has one, and confirms the 2,000 monthly code completions, but the credit figure is not on the pages we checked.
- How long the flex allotment lasts. GitHub says explicitly that it is designed to change. No review date is published.
- Any vendor-published head-to-head benchmark. Neither company has published a like-for-like quality comparison of Claude Code against Copilot’s agent. Every ranking you have read is somebody’s anecdote, including the well-argued ones.
Frequently asked
How is Claude Code different from GitHub Copilot?
Copilot is primarily an in-editor assistant: code completions and next-edit suggestions, unlimited and unbilled on paid plans, with a chat panel, a CLI and a cloud agent that draw on a credit allowance. Claude Code is a terminal agent with no completion feature at all — it reads your repository, runs commands and edits files in place. The billing differs as much as the interface: Copilot publishes an allowance in dollars, Anthropic publishes a five-hour rolling window and a weekly window with no stated size.
Which is cheaper, Claude Code or GitHub Copilot?
By list price, Copilot Pro at $10/month undercuts Claude Pro at $17–20/month, and Copilot tells you the $10 includes $15 of metered usage. Whether it is cheaper in practice depends on whether your work stays inside completions, which cost nothing extra on Copilot, or lives in the agent, which costs credits on Copilot and an unquantified share of your window on Claude. At the $100 tier the published comparison is Copilot Max’s $200 of included credits against Claude Max’s undisclosed 5x or 20x.
Can I use Claude models with GitHub Copilot?
Yes. GitHub’s supported model list includes Claude Sonnet, Claude Opus and Claude Haiku, and paid Copilot plans can select among them. Their tokens are metered against your Copilot credits at published API rates. Copilot Free and Copilot Student reach models only through auto model selection, and organisation administrators can restrict which models their members may use.
Can I run Claude Code with a Copilot subscription?
No. Claude Code requires a Claude Pro, Max, Team or Enterprise plan, or Claude Console access. A GitHub Copilot seat gives you Claude models inside Copilot’s own surfaces, not the Claude Code CLI. Being able to select Opus in Copilot Chat is not the same as running Claude Code, and the two bill to different accounts.
Do unused Copilot AI credits roll over?
No. GitHub’s documentation states that included credits do not carry over between months, that unused credits are forfeited, and that the allowance resets to the full monthly amount at 00:00:00 UTC on the first day of each calendar month. For organisations the pooled allowance resets on the same schedule.
What happens when I run out on each?
On Copilot, individuals set a dollar budget and continue at $0.01 per credit, upgrade, or wait for the monthly reset; code completions keep working regardless. Organisations have additional usage enabled by default, so work continues at published rates unless an administrator turns it off, in which case usage is blocked until the pool refreshes. The old behaviour of falling back to a cheaper model has been removed. On Claude, you wait for the five-hour or weekly window to reset, or enable usage credits — which keeps you working but shortens the prompt cache lifetime from one hour to five minutes.
Which is better for a team of ten?
Neither company publishes evidence that settles it, so decide on what is checkable. Copilot Business at ten seats is $190/month with a pooled 19,000 credits ($190) at the standard rate, plus unlimited completions and four levels of budget control. Claude Team at ten standard seats is $200–250/month with no stated allowance, plus a published rate-limit sizing table if you deploy through the Console instead. If forecastability matters more than the specific agent, Copilot wins on the published record. If your ten developers already prefer Claude Code, no amount of table-reading will change that, and the honest move is a paid pilot with /usage open.
Sources
Every figure above was read from a vendor page in August 2026. Where a page did not contain a number, we said so rather than estimating.
| Source | What we took from it |
|---|---|
| github.com — Copilot Plans & pricing | Plan names and list prices; the monthly credit totals for Pro, Pro+ and Max |
| docs.github.com — Usage-based billing for individuals | 1 credit = $0.01; base and flex allotment split; rollover and reset policy; which features are billed |
| docs.github.com — Usage-based billing for organizations and enterprises | Business and Enterprise per-seat credits; the pooling rule; promotional allowances; budget controls |
| docs.github.com — Plans for GitHub Copilot | Business $19 and Enterprise $39 per granted seat; Copilot Free completions |
| docs.github.com — Supported AI models | Availability of Claude Sonnet, Opus and Haiku inside Copilot; auto-selection limits on Free and Student |
| github.blog — Copilot is moving to usage-based billing | 1 June 2026 effective date; removal of premium request units and fallback experiences |
| claude.com/pricing | Claude plan prices; the multiplier language; the five-hour and weekly window statement |
| code.claude.com — Manage costs effectively | $13/active day, $150–250/month, 90% under $30/day; agent team 7x; cache lifetime; TPM/RPM sizing table |
For the wider field, see our comparison of the coding models themselves, which looks at capability rather than billing.
Update, 1 September 2026: the summer credit subsidy is over, and the seat now returns exactly its price
GitHub’s promotional included usage for existing Copilot Business and Copilot Enterprise customers ran from 1 June to 1 September 2026. It ended today. Because the credit pool resets at 00:00:00 UTC on the first day of each calendar month and unused credits do not carry over, this was not a gradual taper — the lower allowance was in force the moment the month rolled over.
The official figures, from GitHub’s usage-based billing documentation: existing customers received 3,000 AI credits per user per month on Business and 7,000 on Enterprise during the promotional window. The standard amounts, which apply from now on, are 1,900 credits per user on Business and 3,900 on Enterprise. GitHub also fixes the conversion rate: 1 AI credit = $0.01 USD, with usage beyond the pool charged at that same rate.
The line nobody is comparing: the seat price and the included usage are the same number
Put GitHub’s two official numbers next to each other and something falls out that the coverage of this change has mostly skipped. Copilot Business is $19 per seat per month and includes 1,900 credits. At $0.01 per credit, 1,900 credits is worth exactly $19.00. Copilot Enterprise is $39 per seat and includes 3,900 credits — exactly $39.00. (That the two figures coincide is our arithmetic on GitHub’s official list prices and its official credit conversion rate; GitHub does not present it this way.)
So from today, a Copilot seat returns 100% of its price back to you as metered model usage. Everything else the seat carries — the policy and governance controls, audit and spend reporting, the model catalogue, and code completions and next edit suggestions, which GitHub states are not billed in AI credits and remain unlimited on all paid plans — is priced at zero on top of the usage you were going to pay for anyway.
That reframes the summer promotion. It was not a discount on a seat; it was GitHub handing existing customers more model usage than the seat cost. On Business, $30 of included usage on a $19 seat, a ratio of about 1.58x. On Enterprise, $70 on a $39 seat, about 1.79x. (Both ratios are our arithmetic on the official credit counts and the official $0.01 conversion.) As of today both ratios are 1.00x. Nothing was added to your bill — what was withdrawn was a subsidy on the usage side, which is exactly the kind of change that does not show up in a price-comparison table.
What it costs you, per seat and per hundred seats
| Plan | Promotional (Jun–Aug) | Standard (from 1 Sep) | Change in included usage |
|---|---|---|---|
| Copilot Business, $19/seat | 3,000 credits ($30) | 1,900 credits ($19) | −1,100 credits, −$11/seat/mo (−36.7%) |
| Copilot Enterprise, $39/seat | 7,000 credits ($70) | 3,900 credits ($39) | −3,100 credits, −$31/seat/mo (−44.3%) |
Credits pool at the billing entity rather than per user, so the effect scales cleanly. A hundred-seat Copilot Business organisation went from a shared pool of 300,000 credits to 190,000 — a reduction of 110,000 credits, or $1,100 a month of included usage (our arithmetic). Whether that becomes $1,100 of new invoice depends entirely on whether the org was drawing on the promotional headroom. An organisation that never crossed 190,000 credits sees no change at all. One that was routinely running at 250,000 now has 60,000 credits a month of overage at $0.01 each.
The practical instruction is narrow and worth doing this week: pull your August credit consumption before the number is buried, and compare it against the new pool rather than the old one. August was the last month at the inflated allowance, so an August total that sat comfortably inside 300,000 tells you nothing about September.
A second billing change lands on 1 October, thirty days behind this one
GitHub announced on 28 August 2026 that it is reopening Copilot Business and Copilot Enterprise sign-ups for customers paying by credit card or PayPal, with stricter account vetting and revised billing. For new customers, seats must be paid for before users gain access. For existing Business and Enterprise customers on a credit card or PayPal, assigned seats begin incurring an upfront charge at the start of the billing cycle from 1 October 2026. GitHub states plainly that Copilot Business and Enterprise prices are not changing, and that spend controls, usage tracking and the option to buy additional credits all remain.
So two separate changes hit within a month: the included usage dropped today, and the timing of the seat charge shifts on 1 October. Neither is a list-price increase. Both change what shows up on the invoice. If you are budgeting a rollout, they need to be modelled as two events, not one.
Why you cannot line a credit up against a Claude token and call it even
The obvious next move is to convert 1,900 credits into an equivalent volume of Claude API spend and compare. That comparison is less solid than it looks, for a reason that has nothing to do with either vendor’s prices.
Anthropic states in its own pricing documentation that Claude 4.7 and later models use a newer tokenizer which produces approximately 30% more tokens for the same text, with the exact increase depending on content and workload; Claude Sonnet 4.6 and earlier use the previous tokenizer. A per-token or per-credit rate is therefore not a like-for-like unit across vendors, or even across Claude generations. GitHub converts token usage into credits using per-model rates, so routing your Copilot work to a newer Claude model consumes credits against a token count that is not comparable to the one an older model would have produced from the same source file.
We have not found a published cross-vendor tokenisation benchmark we would be willing to stand behind, so we are not going to invent a conversion factor. The honest version is: the credit arithmetic above is exact within GitHub’s own system, and any bridge from it to another vendor’s per-token pricing carries an unquantified error term of roughly this size. Treat headline per-token comparisons between Copilot credits and direct API spend as directional at best. We make the same point about cross-vendor list prices in our breakdown of what AI coding assistants actually cost, and it applies with more force here, because the credit layer adds a second conversion on top of the tokeniser problem.
Does this change the answer on which to buy?
For most teams, no — and it is worth being clear about why, because a 37% cut to included usage sounds like it should move the decision.
It does not, because the seat still returns its own cost in usage. If Copilot Business made sense for you in July at $19 for $30 of credits, the question was never really the credit headroom — it was whether you wanted governance, unlimited completions and a single invoice covering multiple model vendors. All of those survive today’s change untouched. What has changed is that the seat is now priced honestly rather than generously, and teams that had quietly built a workflow on top of the promotional headroom will find out this month.
The teams for whom it does move the answer are the heavy-agent ones: organisations whose credit draw was landing between 190,000 and 300,000 per hundred seats, who were inside the allowance all summer and are now in overage every month. For them the comparison against a metered subscription is worth reopening, and the two pieces to read next are our Codex versus Claude Code comparison, which covers the metered-subscription side of the same decision, and our look at the GLM coding plan against Claude Code, where the same cache-and-tokeniser caveats decide a cost comparison that looks lopsided until you read the individual line items.
Sources for this update, read 1 September 2026: GitHub Docs, “Usage-based billing for organizations and enterprises” (credit amounts, promotional amounts, the $0.01 conversion, pooling and the UTC monthly reset); GitHub Docs, “About billing for GitHub Copilot in organizations and enterprises” (seat prices, unlimited completions); GitHub Changelog, “Upcoming changes to GitHub Copilot policies and billing”, 28 August 2026 (the 1 October billing change); Anthropic, Claude Platform pricing documentation (the tokeniser note). Dollar values, deltas and ratios derived from those figures are our arithmetic and are labelled as such.