Anthropic ships Claude Sonnet 5 as its most agentic Sonnet yet

Anthropic released Claude Sonnet 5 today, and the framing matters more than the version bump suggests. In its launch post, the company calls it 'the most agentic Sonnet model yet,' built to make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models. The implicit comparison is to Sonnet 4.6, which the post positions as the prior baseline.

Pricing is where the announcement has real teeth. Through August 31, 2026 Sonnet 5 launches at $2 per million input tokens and $10 per million output, before stepping to standard rates of $3 and $15. The footnote you actually need to read sits right next to that. Sonnet 5 uses an updated tokenizer, and the same input can map to roughly 1.0 to 1.35 times more tokens depending on the content type. So the headline per-token discount is real but partly eaten by the new accounting.

The safety section is the bit that may matter more than the leaderboard race. Anthropic says Sonnet 5 has an overall lower rate of undesirable behaviors than Sonnet 4.6 and is generally safer to use in agentic contexts. More unusually, the company says it did not deliberately train Sonnet 5 on cybersecurity tasks, and that on evaluations testing potentially dangerous cyber skills such as developing software exploits it shows substantially poorer performance than models like Opus 4.8. That is a product decision, not just a safety footnote.

The honest caveat is that almost every concrete claim sits inside Anthropic's own post. The partner testimonials, including one from Zimu Li, identified as a Member of Technical Staff at an early-access partner, line up with the agentic-execution story but they are early access, not independent. What the announcement does not give you is specific benchmark scores, the model's context window, or how the tokenizer change translates to real bills on production agent workloads.

If the self-verification and tool-use claims hold up under outside testing in the coming weeks, the teams who benefit first are those already building on Claude Code and the Claude Platform, who can lock in the cheaper introductory rates before the September step-up.