Anthropic has released Claude Sonnet 5, the new mid-tier model that becomes the default across all plans (Free, Pro, Max). The news isn’t just about the 1M native token context window, but about agentic capabilities: autonomous planning, browser and terminal use, advanced tool use — features that until recently required Opus 4.8, the largest and most expensive model.

Why this matters for you

If you use Claude for automation or multi-step tasks, the price-to-performance ratio shifts. Sonnet 5 closes complex tasks that Sonnet 4.6 left incomplete: early access testers report it monitors its own output without explicit prompting, completes end-to-end procedures (like updating Salesforce and sending emails) without stopping, and handles “brownfield” codebases — those full of race conditions and hidden tests that nobody wants to touch.

Promotional pricing: $2/M input, $10/M output through August 31 (then $3/$15). But there’s a catch: Sonnet 5 uses a new tokenizer that produces roughly 30% more tokens for the same English text. In early independent tests, the Universal Declaration of Human Rights in English goes from 2,356 to 3,341 tokens (+42%), Python code +28%, Spanish +33%, Mandarin Chinese essentially unchanged. So the effective cost is closer to the old price, except during the promotional period.

If you want to try it

Sonnet 5 is already the default on Claude.ai and available via API under the model ID claude-sonnet-5. Adaptive thinking is enabled by default, so the model plans before acting — if you want to disable it, check the API documentation for the thinking: {type: "disabled"} parameter. The 1M token window and 128K maximum output are native; no prompt engineering needed to make them work.