Claude Sonnet 5 launches with 1M token context window, new tokenizer, and adaptive thinking enabled by default

Anthropic released Claude Sonnet 5, positioning it as having performance close to Opus 4.8 but at lower prices, according to the developer documentation reviewed by Simon Willison. The company also released a system card explaining how the model’s safeguards compare to other releases.

The model introduces a 1 million token context window and supports up to 128,000 maximum output tokens. It retains the same set of tools and platform features as Claude Sonnet 4.6, but discontinues support for sampling parameters temperature, top_p, and top_k.

Adaptive thinking is enabled by default in Sonnet 5, unless developers explicitly disable it via the API. Anthropic’s pricing remains listed at $3 per million input tokens and $15 per million output tokens, with an introductory discount to $2/$10 until August 31.

A key change is the new tokenizer, which produces approximately 30% more tokens for the same input text compared to Claude Sonnet 4.6. This effectively increases costs by the same proportion despite unchanged list prices. Measurements across sample documents show roughly 1.4x more tokens for English, 1.33x for Spanish, 1.28x for Python code, and effectively no change for Simplified Mandarin.

Simon Willison used his Claude Token Counter tool to compare token counts across documents, including the Universal Declaration of Human Rights in English, Spanish, and Simplified Mandarin, as well as a 4,279-line Python file. The results confirm the approximately 30% increase in token counts for most languages tested.