Claude API pricing: every model, every rate

How to read these tables

Every rate above is per million tokens, in US dollars, as Anthropic publishes it for the first-party Claude API, and every one renders from this site's dated facts module rather than being typed onto the page. The same rates apply on the other Anthropic-operated platforms — Claude Platform on AWS and Microsoft Foundry — where the marketplace converts them into consumption units on your AWS or Azure invoice. Amazon Bedrock and Google Cloud set their own prices, which these tables do not cover.

The column that changed most recently is the cache read. Until Fable 5.1 and Opus 5.5, a cache read cost the same fraction of input on every model, so listing input and output was enough. It no longer is: the newest Fable and Opus models read their cache at a deeper discount, which is why a model's cache-read rate is printed beside its headline rate instead of being left for you to derive.

What the tables leave out, on purpose

A per-token rate is not a per-request cost. Three things sit between the two. The tokenizer: models from Opus 4.7 onward count more tokens for the same text than older ones, so identical prompts bill differently at identical rates. Thinking: reasoning tokens bill as output whether or not their text is returned to you, and on several current models thinking cannot be switched off. And tool use: a request that declares tools carries a small, model-specific system prompt, and server tools add their own charges, listed in the last table.

The newest models also bill their entire context window at the standard rate — there is no long-context surcharge on them — and caching and batch discounts stack across the full window.

Turning a rate into a bill

Paste real text into the token & cost estimator to price it on every model at once, use the prompt-caching savings calculator to see what the cache-read column is worth to your request pattern, and the batch vs realtime calculator for work that can wait. Each model's own reference page carries the same rates alongside its limits and lifecycle.