Anthropic has released Claude Sonnet 5, an upgrade to its Claude Sonnet 4.6 model. This new version includes adaptive thinking turned on by default and introduces a new tokenizer. The release also details changes to how certain parameters are handled.
Key Points
- Claude Sonnet 5 is a direct upgrade for Claude Sonnet 4.6.
- Adaptive thinking is now enabled by default in Claude Sonnet 5.
- Manual extended thinking requests now result in a 400 error.
- Setting
temperature,top_p, ortop_kto non-default values also returns a 400 error. - Claude Sonnet 5 supports up to 300k output tokens via the Message Batches API with the
output-300k-2026-03-24beta header. - The model features a 1M token context window by default.
Context
According to Anthropic's documentation, Claude Sonnet 5 represents an evolution in the Sonnet model family. It is designed to be a drop-in replacement for its predecessor, with specific behavioral adjustments. The documentation outlines the model's capabilities, pricing, and lifecycle status.
Why It Matters
Builders using Anthropic's models need to be aware of the behavioral changes in Claude Sonnet 5, particularly regarding adaptive thinking defaults and the handling of sampling parameters. These updates can impact existing integrations and require adjustments to prompting strategies.
What To Do
- Review the documentation for Claude Sonnet 5 to understand all new features and changes.
- Note that adaptive thinking is now on by default and steer its depth using the
effortparameter. - Adjust any code that sets
temperature,top_p, ortop_kto non-default values, as these will now cause a 400 error. - Explore the Message Batches API for workloads requiring up to 300k output tokens.
