Claude API pricing per model, November 2026

The four models at the top of the table

These are the prices at the top of the model pricing table in Anthropic's API documentation as we read it on 13 November 2026. They are in US dollars per 1 million tokens (the page writes MTok).

Output costs five times as much as input on all four models. Claude Haiku 5.5 is the only one priced by prompt length: a request whose prompt is over 100,000 tokens pays the higher row. The other three include the full 1M token context window at the standard price.

What 10 million tokens cost

Per-million prices are hard to picture, so here is the same table multiplied by ten: what 10 million input tokens and 10 million output tokens cost on each model at the standard price, with no caching. A real job is a mix of the two, so read the columns separately and add up your own mix. The Haiku 5.5 row uses the price for prompts up to 100,000 tokens.

Ways to pay less

• Batch API: a 50 percent discount on both input and output tokens for requests processed asynchronously. It does not apply to Claude Managed Agents sessions. Per million tokens that is $4 in and $20 out on Fable 5.2, $2 and $10 on Opus 5.5, $1 and $5 on Sonnet 5.5, and $0.05 and $0.25 on Haiku 5.5 for prompts up to 100,000 tokens.
• Prompt caching: writing a prompt to the cache costs 1.25 times the input price for a 5-minute cache and 2 times for a 1-hour cache. Reading it back costs the cache hit price in the first table. Anthropic says the 5-minute cache saves money after one cache read and the 1-hour cache after two.
• Both together: the page says the caching multipliers stack with the Batch API discount.
• A smaller model: Anthropic's own advice on the page is Haiku for simple tasks, Sonnet for most production workloads and Opus for the most complex reasoning.

What costs more

• Fast mode: up to 2.5 times higher output tokens per second at a premium price. Among the four models above, Fable 5.2 in fast mode is $16 in and $80 out per million tokens and Opus 5.5 is $8 and $40. Fast mode is not available with the Batch API.
• US-only inference: for Claude 4.6 and later models, asking for inference in the United States only multiplies every token price by 1.1. Global routing, the default, is the standard price.
• Long prompts on Haiku 5.5: over 100,000 tokens, the whole request pays the higher row, even when part of the prompt is a cache hit. Prompt length counts all input tokens, including cache reads and cache writes.
• The tokenizer: Anthropic says Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text. The exact increase depends on the content and workload shape. The same text is therefore more tokens on those models than on Claude Sonnet 4.6 or earlier, before the price per token is applied; the price per token in the tables above is a separate matter.

Charges that are not tokens

Tool use is paid in tokens: the tool definitions, the calls and the results count as input and output. A few server-side tools and Claude Managed Agents add their own charge.

More models on the price list

The same table lists more models under Additional models. This is a selection, per million tokens. A lower version number is not a lower price: Claude Opus 5 is $5 and $25 against $4 and $20 for Claude Opus 5.5.

What this article does not cover

• Amazon Bedrock and Google Cloud (Google managed): Anthropic's page says those providers invoice you and points to their own price lists, which we did not read. Claude Platform on AWS and Claude in Microsoft Foundry are rated at the standard per-model prices and billed through the marketplace.
• Volume discounts: the page says they are negotiated case by case, and that enterprise customers can ask for custom pricing.
• Free credits: the page says new users receive free credits to test the API, without an amount.
• Subscriptions are not on this page. Claude Code's plans are covered in our Claude Code pricing article.
• Tax.

Knowing what API work cost per task

A price list tells you the rate. It does not tell you what a task cost, because that depends on how many tokens the work used, how much of the input was cached and which model ran it.

Hourtick does not sell Anthropic's models and does not change what you pay Anthropic. It is a time tracker where an AI agent works on a task and reports its own time and model cost on it, with the model and the tokens, next to the hours people log. The numbers are what the agent reports. If you connect Anthropic, those reports are checked against the provider's bill every month.