Introducing Claude Sonnet 5

Anthropic shipped Claude Sonnet 5 on June 30. The interesting number is not the benchmark. It is the price.

At $2 per million input tokens and $10 per million output tokens through August 31 (then $3 and $15), Sonnet 5 is roughly 1/7th the cost of Opus 4.8. On the knowledge-work benchmark Anthropic used, it slightly outperforms Opus 4.8. That is the sentence I have been waiting for a Sonnet release to earn.

I have spent the last quarter watching teams struggle with the same tradeoff. Opus is smart enough to plan an agentic workflow, but too expensive to run the tail of it. Sonnet 4.6 was cheap enough, but you had to hand-hold it through longer chains. Every team I talked to was building a router, hard call goes to Opus, easy call goes to Sonnet, and the router itself becomes a research project. Sonnet 5 collapses that decision for most product surfaces.

Three numbers matter to anyone shipping an agent.

**SWE-bench Pro: 63.2%.** Sonnet 4.6 was at 58.1%. GPT-5.5 sits at 58.6%. Gemini 3.5 Flash is at 55.1%. The gap over GPT-5.5 is small, but it is a gap, and at a lower price point than any of them.

**Terminal-Bench 2.1: 80.4%.** Up from 67.0% on Sonnet 4.6. GPT-5.5 still edges Sonnet 5 here at 83.4%. This one is closer to a tie than a win, but the delta versus Sonnet's previous ceiling is what tells you about the trend.

**Self-checking behavior.** Anthropic claims Sonnet 5 checks its own output without being asked. I want to see the eval before I take a strong position, but if it holds up, it changes how I design the verifier layer in every agent I have deployed this year.

The system card is the artifact worth reading. It is 145 pages, and it spends most of its budget on how the model behaves inside an agentic loop, prompt injection resistance, hijack recovery, refusal patterns on tool use, rather than benchmark tables. That is exactly the shift I have been asking model cards to make. Benchmarks tell you what a model can do on a good day. System cards should tell you how it fails on the median day.

What this means for teams I advise.

If you are routing between Sonnet and Opus, delete the router. Run Sonnet 5 for the whole loop and only fall back to Opus on the hard 5%. The savings pay for a better eval setup this quarter.

If you are on Sonnet 4.6, plan a two-week migration window. Not because the API changed, it did not, but because your prompts are still tuned to a model that needed more scaffolding. Sonnet 5 wants less handholding. Give it less.

If you are on a non-Anthropic model for cost reasons, re-run your evals. At $2 and $10 through August, the cost argument that pushed you off Anthropic six months ago may not survive contact with a fresh benchmark. I am not saying switch. I am saying measure.

One caveat on all of this. I have not personally shipped Sonnet 5 into a production agent yet. Everything above is from the announcement, the system card, and half a day of API testing. If you find a place where the model quietly regresses versus Sonnet 4.6, reply with the trace. I will update this post publicly.

Two questions I am chewing on.

The first: how much of the agentic gain is the model, and how much is Anthropic's tool-use scaffolding around it? Sonnet 5's browser and terminal skills are impressive because they are jointly tuned with Anthropic's own agent harnesses. Teams that have built their own harness may see less of the lift. That is worth a measured comparison, not a hot take.

The second: what does the pricing curve mean for the small-model-beats-big-model thesis I keep coming back to? A 7B vertical model at 1/40th the cost of a frontier call still beats Sonnet 5 on the narrow tasks I care about, legal extraction, medical chart parsing, compliance flagging. But the horizon at which that argument fails just moved. If Sonnet 5 is good enough on your vertical, you may not need the small model after all. Worth an eval, not a hot take.

I will spend the next week running Sonnet 5 against three portfolio companies' production loops, one legal, one health, one household ops. If you want the results before I write them up in the next Brief, reply and I will share.