Read on 13 October 2026, per million tokens in US dollars: Claude Fable 5.1 costs $10 input and $50 output, Opus 5.5 $4 and $20, Sonnet 5.5 $2 and $10, and Haiku 5.5 $0.10 and $0.50 on prompts up to 100,000 tokens. The Batch API takes 50% off, and reading cached input costs 2.5% to 10% of the input rate. To estimate a bill, count input and output tokens separately, multiply each by its rate, then apply batch, caching and long-prompt rules.
Claude API pricing is per million tokens: on 13 October 2026, Fable 5.1 costs $10 input and $50 output, Opus 5.5 costs $4 and $20, Sonnet 5.5 costs $2 and $10, and Haiku 5.5 starts at $0.10 and $0.50. Batch jobs are half price, cached input costs a small fraction of the input rate, and a chat subscription does not pay for any of it.
Every rate here was read on 13 October 2026 from Anthropic's pricing docs, which state that all prices are in US dollars, with details from the models overview and the prompt caching, batch processing and token counting guides. The worked examples use made-up workloads and the real rates. This post contains no speed or quality benchmarks, because Anthropic's pricing pages do not publish any.
Input costs a fifth of output on every model: $10, $4, $2 and $0.10. Source: Anthropic pricing docs, read 13 October 2026
The full rate table
A token is a piece of text, and a large language model (see LLM in the glossary) is billed on tokens in and tokens out. Output always costs five times input. Prices per million tokens:
| Model | Input | 5-minute cache write | 1-hour cache write | Cache read | Output |
|---|---|---|---|---|---|
| Claude Fable 5.1 | $10 | $12.50 | $20 | $0.25 | $50 |
| Claude Opus 5.5 | $4 | $5 | $8 | $0.20 | $20 |
| Claude Sonnet 5.5 | $2 | $2.50 | $4 | $0.10 | $10 |
| Claude Haiku 5.5, prompts up to 100,000 tokens | $0.10 | $0.125 | $0.20 | $0.01 | $0.50 |
| Claude Haiku 5.5, prompts over 100,000 tokens | $0.50 | $0.625 | $1 | $0.05 | $2.50 |
All four models have a 1M-token context window and can write up to 128K tokens in one reply. Older models are still sold at their own rates, for example Claude Opus 5 at $5 and $25 and Claude Haiku 4.5 at $1 and $5, so check which model ID your code calls before you assume the new prices apply.
How to estimate a bill
- 1Count input tokens per request
System prompt, tool definitions, documents and the conversation so far. The token counting endpoint is free and returns an estimate.
- 2Estimate output tokens per request
Read the usage field of a few real responses. Output is the expensive side.
- 3Multiply by volume
Requests per month times tokens per request, input and output separately, in millions.
- 4Multiply by the rate
Input millions times the input price plus output millions times the output price.
- 5Apply the modifiers
Halve it for batch, re-price the cached part, and check the long-prompt rule for your model.
Five worked examples. The workloads are illustrative; the rates are the ones in the table. To run the same math on your own prompt, the free AI token cost calculator takes pasted text or token counts and shows the cost per request, per 1,000 and per month.
| Workload | Tokens | Math | Cost |
|---|---|---|---|
| 10,000 support replies on Haiku 5.5 | 3,000 in and 400 out each | 30M x $0.10 + 4M x $0.50 | $5.00 |
| 1,000 document summaries on Sonnet 5.5 | 20,000 in and 1,000 out each | 20M x $2 + 1M x $10 | $50.00 |
| The same summaries through the Batch API | As above, results can wait | Half of $50.00 | $25.00 |
| A 10,000-token system prompt on Opus 5.5, sent 1,000 times | Uncached | 10M x $4 | $40.00 |
| The same prompt with a 5-minute cache | One write, 999 reads | 10,000 x $5 per million + 9.99M x $0.20 | $2.05 |
Two things stand out. Model choice moves the bill more than anything else: the support example costs $5 on Haiku 5.5 and would cost $200 on Opus 5.5 for the same tokens. And caching only pays when requests keep arriving before the cache expires. The $2.05 figure assumes all 999 reads land inside the five-minute lifetime, which each read refreshes.
Discounts and surcharges
| Modifier | Effect on price | Applies to |
|---|---|---|
| Batch API | 50% off input and output | Work that can wait. Most batches finish within an hour |
| 5-minute cache write | 1.25x the input price | A prefix reused within five minutes |
| 1-hour cache write | 2x the input price | A prefix reused across a longer session |
| Cache read | 0.1x input. 0.05x on Opus 5.5 and Sonnet 5.5, 0.025x on Fable 5.1 | Every reuse of a cached prefix |
| US-only inference | 1.1x on every token category | Workloads that must run in the US |
| Fast mode on Opus 5.5 | $8 input and $40 output | Research preview, Claude API only, no batch |
| Web search tool | $10 per 1,000 searches, plus tokens | Each search, whatever it returns |
These stack. Anthropic says the caching multipliers combine with the batch discount and with data residency pricing. A few details from the guides are worth knowing before you plan around them:
- Batch. A batch holds up to 100,000 requests or 256 MB. Most finish within an hour, results are available when every message is done or after 24 hours, and a batch that has not finished by then expires (batch processing).
- Caching. The simplest route is automatic caching: one
cache_controlfield at the top level of the request. The 5-minute cache pays for itself after one read and the 1-hour cache after two (prompt caching). - Long prompts. Fable 5.1, Opus 5.5 and Sonnet 5.5 bill a 900,000-token request at the same per-token rate as a 9,000-token one. Only Haiku 5.5 has a step.
- Web fetch has no extra charge, and code execution is free when used with web search or web fetch. On its own it is billed by container time after 1,550 free hours a month per organization.
Token costs people miss
- The tokenizer. Models from Claude Opus 4.7 onward use a tokenizer that produces about 30% more tokens for the same text. A lower price per token is not automatically a lower price per task when you migrate.
- Tool definitions. Tools add input tokens: the tool names, descriptions and schemas, plus a tool use system prompt of 286 tokens on Opus 5.5, Sonnet 5.5 and Haiku 5.5.
- Effort. The
effortsetting controls how many tokens a model spends. Opus 5.5 and Haiku 5.5 default tomedium, Fable 5.1 and Sonnet 5.5 tohigh, and lower effort saves tokens at some cost in capability (effort). - Search results. Web search results are counted as input tokens, in the turn that fetched them and in later turns.
Tools are where this adds up fastest. Our Claude tool use guide shows what a tool definition looks like, and the Claude API tutorial shows where the usage numbers appear in a response.
Which model to start on
Anthropic's models overview says to start with Opus 5.5 for most workloads and to use Fable 5.1 when your evaluations on Opus at higher effort still fall short. Its pricing page gives the cost-side rule: Haiku for simple tasks, Sonnet for most production workloads, Opus for the most complex reasoning.
In practice that means routing. Send classification, extraction and routing to Haiku 5.5, drafting and everyday generation to Sonnet 5.5, and keep Opus 5.5 or Fable 5.1 for the steps that fail on a smaller model. Decide with your own test set, not with a table: run the same 20 real inputs through two models and compare the output and the usage numbers.
Claude API vs OpenAI API pricing
Both companies' price lists were read the same day, Anthropic's from its pricing docs and OpenAI's from its API pricing page. Prices are input and output per million tokens.
| Tier | Claude | OpenAI |
|---|---|---|
| Top | Fable 5.1: $10 and $50 | GPT-6 Astra: $10 and $50 |
| Middle | Sonnet 5.5: $2 and $10. Opus 5.5: $4 and $20 | GPT-6.1 Sol: $2 and $10 |
| Bottom | Haiku 5.5: $0.10 and $0.50 | GPT-6 Luna: $0.10 and $0.50 |
| Long prompts | Standard rate to 1M tokens, except Haiku 5.5 above 100,000 | 2x input and 1.5x output above 272K input tokens |
| Cached input, top model | $0.25 | $1.00 |
| Batch | 50% off | 50% off, on Batch and Flex |
List price does not settle it, because the tokenizers differ and so does how much each model writes. For a wider view, see the LLM API pricing comparison, Claude vs GPT and Claude vs Gemini. The app plans are compared in Claude vs ChatGPT.
Credits, caps and free usage
The API runs on prepaid credits bought in the Claude Console. Credits expire a year after purchase and are non-refundable, and you are billed only for successful calls (help centre). Each usage tier has a monthly spend cap, $500 on Start, $1,000 on Build and $200,000 on Scale, and you can set a lower limit of your own (rate limits).
New users get a small amount of free credits to test with; Anthropic gives no figure. Setup is covered in how to get a Claude API key, and the difference between the API and the Pro, Max and Team plans is in the Claude pricing guide. If you are pricing a product of your own on top of these rates, AI SaaS Builder, one of our programs, has a lesson on cost optimization and caching inside its Claude API module. More guides are in the AI automation library.
What changes often
Models and prices changed several times in the six weeks before this was written. Treat the numbers above as a reading, and check the pricing docs before you quote a client or set a budget.
- 1 Sep 2026Claude Sonnet 5 stays at $2 and $10
A rise to $3 and $15 scheduled for this date did not happen. The launch price is now the standard price.
- 22 Sep 2026Claude Opus 5.5 launches at $4 and $20
Claude Opus 5 is listed at $5 and $25.
- 28 Sep 2026Claude Sonnet 5.5 launches at $2 and $10
- 7 Oct 2026Claude Haiku 5.5 launches from $0.10 and $0.50
Sonnet 5.5 cache reads drop from $0.20 to $0.10 the same day.
- 30 Nov 2026Claude Sonnet 4.5 retires from the Claude API
Scheduled. Anthropic recommends moving to Sonnet 5.5.
Source: Anthropic release notes, pricing docs and deprecations page, read 13 October 2026
Dates are from Anthropic's release notes and model deprecations page.
Claude API pricing: FAQ
How much does the Claude API cost?
Per million tokens on 13 October 2026: Claude Fable 5.1 costs $10 input and $50 output, Opus 5.5 costs $4 and $20, Sonnet 5.5 costs $2 and $10, and Haiku 5.5 costs $0.10 and $0.50 on prompts up to 100,000 tokens. The Batch API halves those prices, and reading cached input costs between 2.5% and 10% of the input rate.
Is the Claude API free?
No. Anthropic's pricing docs say new users receive a small amount of free credits to test the API and give no figure. After that you buy prepaid credits and pay per token. Counting tokens before you send a request is free, and so is the web fetch tool beyond its token cost.
Is the Claude API cheaper than OpenAI's?
At list price the current lineups match tier for tier: $10 and $50 at the top, $2 and $10 in the middle, $0.10 and $0.50 at the bottom. The rules differ. OpenAI doubles the input price once a prompt passes 272K input tokens, while Claude bills most models at the standard rate across 1M tokens. Claude Haiku 5.5 has its own step at 100,000 tokens.
How does prompt caching reduce Claude API cost?
You pay a premium once to store a prompt prefix, then a fraction of the input price each time it is reused. A 5-minute cache write costs 1.25x the input rate, and a read costs 0.1x on most models, 0.05x on Opus 5.5 and Sonnet 5.5 and 0.025x on Fable 5.1. Anthropic says the 5-minute cache pays for itself after one read.
What is the cheapest Claude model?
Claude Haiku 5.5, launched on 7 October 2026, at $0.10 input and $0.50 output per million tokens on prompts up to 100,000 tokens. Above that the whole request costs $0.50 and $2.50. Anthropic describes it as the model for high-volume, latency-sensitive tasks such as classification, extraction and routing.
Does a Claude Pro subscription cover API usage?
No. Anthropic's help centre says the Pro plan does not include API usage, which is billed separately through the Claude Console. Max and Team plans now include monthly API credits, $100 on Max 5x and $200 on Max 20x, claimed by linking a Console organization.
How do I estimate tokens before I send a request?
Use the token counting endpoint with the model ID you plan to call. It is free and returns an estimate. As a rough guide, Anthropic says 1M tokens is about 555,000 words on the current tokenizer, which works out to roughly 1.8 tokens per English word. Models from Claude Opus 4.7 onward use this tokenizer, which produces about 30% more tokens than earlier models for the same text.