Anthropic launches Claude Sonnet 5.5 across coding tools

BRO… 4 out of 5 smartest models on the AA Intelligence Index are now Claude. and Sonnet 5.5 is scored HIGHER than Fable 5.1. whatever Fable version they’re distilling into Opus and Sonnet, is absurd.

The price of frontier-grade AI just dropped about 90% in 90 days. Sonnet 5 launched June 30 as Anthropic's best price-performance model. Today's Sonnet 5.5 beats its top scores on several benchmarks while spending roughly a tenth of the money per task. Same $2 per million token list price. The savings come from somewhere sneakier. Your AI bill is price times tokens consumed, and everyone watches the wrong half. List prices barely moved all year. Token efficiency is where the collapse is happening. Sonnet 5.5 needs far fewer tokens to finish the same work and hits Sonnet 5's best scores at lower effort settings. Output speed jumped 30% on top. Here's what should worry the flagship tier. Opus 5.5 shipped six days ago as the premium model, and Sonnet 5.5 already lands within 2 points of it…

Claude Sonnet 5.5 is now available in Devin Desktop and Devin CLI. On FrontierCode 1.1 Main, it scores 64.4%, a significant improvement from Sonnet 5 (56.2%), surpassing Fable 5.1 at extra high reasoning effort.

The most important thing about Sonnet 5.5 is that it's now the model that powers the free tier on - so all of this stuff can be done by free users ChatGPT's free tier is still GPT-5.6 Luna, which is a lot less capable

Sonnet 5.5 built a working clone of Proof, our open-source document editor, from scratch. It did it at low effort. @kieranklaassen uses this as a test for coding models. Before Sonnet 5.5, only three models he’d tested had managed it: • Fable 5 • GPT-6 Astra • Opus 5.5 Opus 5 couldn’t do it. Sonnet 5.5 joined the group without needing a higher reasoning setting. The result wasn’t as strong as Opus 5.5’s, and a working clone isn’t a claim that it’s ready for production. Kieran still reaches for Opus for longer, more detailed builds. The build, plus our team’s coding, design, and writing tests:

Same day delivery: Sonnet 5.5 is now in Notion.

Turning up Sonnet 5.5’s effort setting didn’t improve its score on our coding test. @kieranklaassen ran the same 15 tasks at three settings: • Low: 9 passed • Medium: 9 passed • High: 9 passed The setting controls how much reasoning the model does. In this test, asking it to think harder didn’t get more tasks right. For design work, @TylerNishida is also happy to stay at low or medium. The design pieces in this video were made at medium effort. Tyler got strong results at higher settings, too. His recommendation: “Max and ultracode effort created incredible results, too, but at that point, you might as well reach for Opus 5.5.” Our starting point: Use medium for work you can review and steer. Reach for Opus when the job calls for a longer, more detailed build. Full Sonnet 5.5 Vibe Check:

.@TylerNishida asked Sonnet 5.5 for video ideas. It made finished videos. That wasn’t the only time it treated a request to think as permission to act: • Asked for video ideas → made the videos. • Asked what part of a visualization meant → built another simulation without answering the question. • Asked to use one skill, a reusable set of instructions → loaded several at high effort. Sometimes that eagerness gets you further than you expected. It can also leave you reviewing work you never requested, before you’ve even decided what to make. Tell it where the brief ends. If you’re exploring ideas, say: “Ideas only. Don’t build anything yet.” If you want an explanation, ask it to explain the existing work before changing it. What we learned from a week with Sonnet 5.5:

Sonnet 5.5 is an incredible model. You probably shouldn't use it. Here's a video explaining why.