Simon Willison
★ 9/10 signal
What's new in Claude Sonnet 5
reasoningcontext
What happened
Anthropic has released Claude Sonnet 5. Key technical changes include: 1) Sampling parameters (temperature, top_p, top_k) are no longer supported. 2) A massive 1 million input and 128k maximum output token context window. 3) "Adaptive thinking" (reasoning) is enabled by default. 4) A new tokenizer that generates ~30% more tokens for the same English, Spanish, and Python text, resulting in an effective 30% price increase despite nominal pricing remaining at $3/$15 per million tokens.
Why it matters
Claude Sonnet 5 introduces a 128k output limit and default reasoning, but deprecates traditional sampling parameters and introduces a new tokenizer that increases effective costs by ~30%.
The take
The deprecation of temperature and top_p is a massive shift, signaling that reasoning models are moving toward deterministic, internal-search-based generation where user-defined sampling harms performance. The 128k output limit is a game-changer for long-form generation and agentic code-writing. However, the 30% "tokenizer tax" is a sneaky price hike that developers must factor into their unit economics immediately.
Do this
Update your API integration code to remove temperature, top_p, and top_k parameters when calling Sonnet 5, and recalculate your token budget/pricing models to account for the 30% tokenizer overhead.
Don't read this site daily. Get it in your inbox.
The daily brief and Sunday deep dive — distilled, scored, and opinionated. For builders only.