What's new in Claude Sonnet 5
Anthropic released Claude Sonnet 5, which performs close to Opus 4.8 at lower prices. Key changes include a new tokenizer producing ~30% more tokens (effectively a 30% price increase), a 1M token context window, removal of temperature/top_p/top_k sampling parameters, and adaptive thinking enabled by default.
Background
- Anthropic is the AI company behind the Claude family of large language models (LLMs). Sonnet, Opus, and Mythos are model tiers: Sonnet is their mid-range "workhorse" offering, Opus is the high-end flagship, and Mythos is a top-secret model so capable in cybersecurity that US government regulations restrict its release.
- This post covers Claude Sonnet 5, the latest mid-tier model. The key developer angle: Anthropic managed to get regulatory approval by proving Sonnet 5 is weaker at cyber-attack tasks than Mythos 5, even though its general performance rivals the older Opus 4.8 flagship.
- The biggest practical change for developers is a new tokenizer — the system that chops text into tokens (the units models process and bill by). Sonnet 5 uses roughly 30% more tokens for the same English text compared to Sonnet 4.6, which means an effective 30% price hike despite unchanged per-token prices.
- Other notable changes: sampling parameters (temperature, top_p, top_k) are no longer configurable; "adaptive thinking" (the model's internal reasoning allocation) is now on by default; and output limits have jumped to 128,000 tokens, enabling generation of very long documents or code.