Messages API

Everything on the Developer Platform runs through the Messages API. You send messages and a model ID; you get a response. Tools, caching, batching, thinking and structured output are all parameters and headers on that one call.

Two parameter changes that break old code

temperature, top_p and top_k are deprecated from Claude Opus 4.7 onward. They still exist in the SDK request types, so your code type-checks — and then returns a 400 error at runtime if you set them to a non-default value. Remove them and steer with prompting.

Thinking works differently per model generation. There is no single correct way to write this across models:

Model	thinking.type: "enabled"	Adaptive thinking
Fable 5	No	Yes, always on
Opus 5	No	Yes
Sonnet 5	No	Yes
Haiku 4.5	Yes	No
Sonnet 4.5, Opus 4.5	Yes	No

Code targeting both Haiku 4.5 and a Claude 5 model has to branch.

Effort

On Opus 5 and Sonnet 5 the effort parameter defaults to high on the Claude API and Claude Code. On Opus 4.8 it defaults to high everywhere, including claude.ai. Set it explicitly if you want a different level — this is a meaningful cost lever.

Extended output

Max output is 128K tokens on current models via the synchronous API. On the Batch API, Opus 5, Opus 4.8, 4.7, 4.6, Sonnet 5 and Sonnet 4.6 support up to 300K output tokens with the output-300k-2026-03-24 beta header.

Discover capabilities instead of hard-coding them

The Models API lists every available model with max_input_tokens, max_tokens and a capabilities object. Anything that needs to survive a model change should read that rather than a table someone pasted in last year.

What it is not

• Not the Claude apps. A claude.ai subscription is not API credit and does not share limits. See API vs the Claude apps.
• Not the Agent SDK. The SDK builds the agent loop on top of this.