Claude API pricing: what every model costs and how to pay less

Claude API pricing is pay as you go, in US dollars per million tokens, and three choices decide most of the bill: the model, whether the start of your prompt can be cached, and whether the work can wait for a batch. On October 8, 2026, Anthropic's price list runs from $0.10 per million input tokens and $0.50 per million output tokens on the new Claude Haiku 5.5 to $10 and $50 on Claude Fable 5.1. This guide lists every current price from Anthropic's pricing page, explains the rules that move the bill (caching, batch, long prompts, thinking, tools and US-only processing) and works through what one real workload costs on each model.

Claude API prices per model

Table: Claude API prices per million tokens (USD, checked October 8, 2026)

A one-hour cache write costs twice the input price ($20 on Fable 5.1, $8 on Opus 5.5, $4 on Sonnet 5.5). Four rules sit behind the table:

• Output costs five times input on every model listed, so long answers weigh more than long prompts.
• Cache reads are not a flat 10%. Anthropic prices them at 0.1 times the input price by default, 0.05 times on Opus 5.5 and Sonnet 5.5 and 0.025 times on Fable 5.1. That is why a Sonnet 5.5 cache read costs $0.10 and not $0.20.
• Haiku 5.5 is the only current model priced by prompt length. A request with more than 100,000 input tokens pays the higher row. Every other current model bills its full 1 million token context window at the same rate: Anthropic's example is that a 900,000-token request costs the same per token as a 9,000-token one.
• Older models are still on sale. Claude Opus 5 and Opus 4.5 to 4.8 cost $5 input and $25 output, Claude Sonnet 5 costs $2 and $10 with cache reads at $0.20, and Sonnet 4.6 and 4.5 cost $3 and $15. Claude Mythos 5.1 has Fable 5.1's prices but is only open to organizations verified through Anthropic's access programs.

Anthropic describes Haiku 5.5 as the model for "high-volume, latency-sensitive tasks such as classification, extraction, and routing", Sonnet 5.5 as "the best combination of speed and intelligence" and Fable 5.1 as the one for "demanding reasoning and long-horizon agentic work". All four current models have a 1 million token context window and up to 128,000 output tokens per request. If you are new to the unit behind these prices, start with what a token is and how a context window works.

What changes the bill beyond the price per token

Thinking is billed as output. On Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, adaptive thinking is on by default, and Anthropic's documentation says "the tokens Claude spends reasoning are billed as output tokens, even when the thinking text isn't returned to you". The thinking text is hidden by default on these models, but it is billed the same. On Opus 5.5 and Fable 5.1 thinking cannot be switched off (the API returns an error); on Haiku 5.5 it can be turned off at high effort or below. You steer how deeply the model thinks with the effort setting, whose default is medium on Opus 5.5 and Haiku 5.5 and high on Sonnet 5.5 and Fable 5.1. For a cost estimate, count the thinking in your output tokens, and measure it on your own prompts.

Prompt caching. Writing a prompt prefix to the cache costs 1.25 times the input price for five minutes or 2 times for one hour; reading it costs the cache-read price in the table. Anthropic says caching pays off after one read with the five-minute cache and after two with the one-hour cache. The prefix must be at least 512 tokens on the current models (4,096 on Haiku 4.5); shorter prompts are processed without caching and without an error. The five-minute lifetime is refreshed at no cost each time the cache is read. Our prompt caching guide explains how to order a prompt so that it gets hits.

Batch API. Requests that can wait for an asynchronous result cost 50% less on both input and output, and Anthropic says the batch and caching discounts stack. Batch is not available with fast mode.

Fast mode, US-only processing and tools.

• Fast mode gives faster output on Opus 5.5 for $8 input and $40 output per million tokens, twice the normal price. Anthropic now describes it as a research preview only on Opus 5 and Opus 4.8.
• Asking for US-only inference (inference_geo: "us") multiplies every token price by 1.1 on Claude 4.6 and later models. The default global routing uses the standard prices.
• Web search costs $10 per 1,000 searches plus the tokens of the results. Web fetch has no extra charge beyond the tokens it brings in.
• Code execution is free when the request also includes the current version of web search or web fetch. Otherwise each organization gets 1,550 free hours a month and then pays $0.05 per hour per container.
• Any request with tools adds a hidden tool-use system prompt: 286 tokens on Opus 5.5, Sonnet 5.5 and Haiku 5.5, plus the tokens of your own tool definitions.

Why Claude counts more tokens for the same text

Anthropic documents that Claude 4.7 and later models, which includes every current model except Haiku 4.5, use a newer tokenizer that "produces approximately 30% more tokens for the same text". Two practical consequences:

• Compare the cost of a task, not the price per token. A prompt measured on Haiku 4.5 or on another vendor's model will count more tokens on Sonnet 5.5.
• Measure before you budget. Anthropic's token counting endpoint returns the exact count for a request on a given model, and every response reports input, output, cache-write and cache-read tokens in its usage data.

What a real workload costs on each model

Take a support assistant that answers 100,000 questions a month. Each request sends 2,000 tokens of fixed instructions and reference text that can be cached, plus 1,000 new tokens (the question and recent messages), and gets back 400 tokens including thinking. Token counts are on the current tokenizer.

On Claude Sonnet 5.5 the month costs:

• New input: 100 million tokens × $2 per million = $200
• Cached input: 200 million tokens × $0.10 per million = $20
• Output: 40 million tokens × $10 per million = $400
• Total: $620 a month, against $1,000 if nothing were cached ($600 of input plus $400 of output).

Table: monthly cost of the sample workload (our calculation from Anthropic's prices, October 8, 2026)

The Batch column only applies to work that can wait, such as classifying yesterday's tickets overnight; a live chat cannot use it. The figures leave out cache writes (with steady traffic the five-minute cache is refreshed on every hit, so writes only happen after quiet gaps), web search, US-only processing and tax.

Prompt length matters on Haiku 5.5. Summarizing a 90,000-token document with a 2,000-token answer costs about $0.01 (90,000 × $0.10 plus 2,000 × $0.50, per million). The same summary of a 150,000-token document costs about $0.08, because the whole request moves to the over-100,000 row. Splitting a long document into chunks under the threshold can cut the price per token by five, at the cost of more requests.

How to choose a model and keep the bill under control

1. Start with the cheapest model that passes your own tests. Run 50 to 100 real requests on Haiku 5.5 and Sonnet 5.5, compare quality and token counts, and only move up to Opus 5.5 or Fable 5.1 where the cheaper model fails.
2. Cache everything that repeats. Put fixed instructions, tool definitions and reference documents first, and the changing part last.
3. Send work that can wait to the Batch API. Reports, classification and evaluations rarely need an instant answer.
4. Watch output and thinking. Ask for short answers, set max_tokens, and lower the effort on simple tasks.
5. Set limits. Anthropic's Console tracks your usage, and your usage tier (Start, Build or Scale) sets your rate limits. New accounts get a small amount of free credit to test.

If you use Claude for yourself rather than in your own software, a subscription may cost less than the API; we compare the two, with Claude Code, in our guide to Claude plans and prices, and how to use Claude Code covers the setup. For a side by side with OpenAI, Google and DeepSeek, see our comparison of LLM API prices, last checked on October 6, 2026.