Calculator
Price a prompt, or see what a budget buys
Per-token prices are hard to feel. Paste text and this page estimates what it will cost. Put in a token count and it prices it. Put in a budget and it tells you how far that budget goes.
Estimate tokens from text
Counting at roughly four characters per token for English. That is an estimate, not a count — we don't ship Anthropic's tokenizer, and a character rule is not one. Code, logs and non-English text break the rule in both directions. Nothing you paste here leaves your browser.
Price a token count
The support-bot preset — 2,000 input, 500 output — costs $0.009 per request on Sonnet 5.5 and $0.018 on Opus 5.5. Both figures come from the rate card below, so they move when it moves.
Reverse a budget
Formula, everywhere on this page: tokens ÷ 1,000,000 × the price per million tokens. Everything here is list price, and it excludes caching, batch discounts and any negotiated rate.
Rate card
How Claude token pricing works
Anthropic bills Claude by the token, and a token is not a word. It is a chunk of text, usually a few characters long. Every exchange is billed twice: once for what you send, once for what comes back. Output costs five times input on both models, which is why a paragraph the model writes costs more than a paragraph you paste in.
The card below is the only price list on this site. The calculator above reads it directly, so the two cannot drift apart: change a number in the table and the estimate follows, with no second copy to keep in step. Haiku 5.5 has been announced but not priced, so its row stays blank on purpose — a blank cell is honest, a guessed one is not.
Three things move a bill more than people expect. Volume, because every price here is per million tokens. Output length, because that is the expensive side. And caching, which has its own section further down.
| Model | Input | Output | Cache read | Cache write (5 min) |
|---|---|---|---|---|
| Claude Sonnet 5.5 checked 2026-09-30 |
$2.00 | $10.00 | $0.20 | $2.50 |
| Claude Opus 5.5 checked 2026-09-30 |
$4.00 | $20.00 | $0.20 | $5.00 |
| Claude Haiku 5.5 | — | — | — | — |
Rates above are US dollars per million tokens, billed on Anthropic's direct API. Last verified September 30, 2026.
By the number
Claude token cost, by amount
Nobody budgets in millions of tokens, so here are the round numbers, priced at both rates for input and output separately.
| Tokens | Sonnet 5.5 input | Sonnet 5.5 output | Opus 5.5 input | Opus 5.5 output |
|---|---|---|---|---|
| 1,000 | $0.002 | $0.01 | $0.004 | $0.02 |
| 10,000 | $0.02 | $0.10 | $0.04 | $0.20 |
| 100,000 | $0.20 | $1.00 | $0.40 | $2.00 |
| 200,000 | $0.40 | $2.00 | $0.80 | $4.00 |
| 500,000 | $1.00 | $5.00 | $2.00 | $10.00 |
| 1,000,000 | $2.00 | $10.00 | $4.00 | $20.00 |
Each figure is the published rate divided by a million and multiplied by the token count — the same arithmetic the calculator runs. These are uncached input and output; a stable prefix you resend is cheaper, which the caching section covers. Note the shape of the numbers: output is five times input, so a long answer costs more than a long question.
Caching
How caching changes Claude token cost
Prompt caching is the one lever that changes a bill without changing the work. Two prices do all of it. Writing a prefix into the cache costs 25% more than sending it as ordinary input: $2.50 against $2.00 on Sonnet 5.5, $5.00 against $4.00 on Opus 5.5. Reading it back costs $0.20 on both — a tenth of Sonnet's input rate, a twentieth of Opus's.
The trade pays only on repetition. Send a prefix once and you have paid the write premium and collected nothing on it. Send it twice and you are already ahead. Send it ten times and the input line costs about a fifth of what it would have: $4.30 for ten uses of a one-million-token prefix, against $20.00 uncached. The saving climbs towards ten times the longer you repeat, but ten times is the ceiling — reaching it means reusing the same prefix more or less forever.
Two details decide whether it works for you. Cache entries expire, and the write price above is for the five-minute cache. A prefix that shows up once an hour can expire between uses, and then you have paid the premium for nothing. And the premium lands on write, so a workload that sends a different document every time does worse than one that never turned caching on. The workloads that win are dull: a fixed system prompt, fixed instructions, the same reference material on every call.
Counting
Counting Claude tokens: what we can and can't do
We estimate token counts; we don't count them. The difference matters, so the box at the top of this page says it plainly rather than burying it in a footnote.
The rule of thumb is four characters per token for English prose. It holds up reasonably well on ordinary sentences and falls apart on everything else. Code, JSON and logs usually run fewer characters per token, because punctuation and short identifiers split into more pieces than plain words do. Accented and character-based languages run the other way, well over four. A mixed document can be wrong in both directions at once.
When the exact number matters — sizing a batch job, or checking whether a prompt fits inside a context window — call Anthropic's count_tokens endpoint against the same model you will actually use. It returns the real count for the real text, and it is free to call. Treat our estimate as a budgeting aid, not as a limit check.
FAQ
Claude token cost: common questions
Method
How we check these rates
Every figure in the card above was read from Anthropic's own announcement page for that model, and the link inside each cell goes to the page the number came from — not a summary, not an aggregator, not somebody else's cached table. Each priced row carries its own check date, and the note under the table carries the date of the last full audit.
Where we cannot source a figure, the cell stays empty. Haiku 5.5 is the current case: announced, not priced. Prices move when a model ships or is re-priced, not on a schedule, so we re-read on announcements and audit the full table monthly.
This page covers cost only. Model choice, context windows and the wider billing picture are on our home page.