Claude Sonnet 5.5 Is Out: The 10 Things That Matter

Claude Sonnet 5.5 came out today, September 28, 2026. Same price as Sonnet 5. Over 30% faster. And on a test of real jobs done alone, it went from 10.3% to 70.6%.

It's the second model in the Claude 5.5 family. Opus 5.5 came out September 22. Haiku 5.5 is coming in the next few weeks. I went through Anthropic's announcement and pulled out the 10 things that matter if you build with Claude Code.

Sonnet 5.5 writes its answers over 30% faster than Sonnet 5. It also needs far fewer tokens to finish a job, so each task costs up to 30% less.

Testers said it batches steps together, so it takes fewer of them. Balyasny ran it on 2,441 finance tasks. Sonnet 5.5 used about 121,000 tokens per answer. Sonnet 5 used 497,000. Zendesk got its support tickets processed 20% faster.

Speed is easier to feel than to read about. The interactive page runs the old and new Sonnet side by side so you can watch the gap. Open the interactive page.

The price didn't move. Sonnet 5.5 is $2 per million input tokens, $10 per million output tokens, and $0.20 per million for cache reads. That's the same as Sonnet 5.

Opus 5.5 is $4 in and $20 out, twice the price. So you get a faster, smarter model for the same money, and it uses fewer tokens per job on top of that.

This is the number I care about most. Terminal-Bench 4.0 gives the AI a real job and lets it work alone. It only counts if the result works at the end.

Sonnet 5.5 scored 70.6%. Sonnet 5 scored 10.3%. Opus 5.5 scored 66.4%. Those are each model's best setting, where they cost about the same per job. Even on the default Medium effort, Sonnet 5.5 beats Sonnet 5's best score for less than a tenth of the cost per task.

Here's what those jobs look like:

Companies saw the same thing. Base44 ran 118 real app builds. Sonnet 5.5 needed 3.6 tries per build. Opus 5 needed 7.7. It rarely stopped to ask the user questions. Box said it rechecks data in the source documents and catches errors Sonnet 5 missed. Testers also said it gets up to speed on big existing projects fast and handles tasks that run for hours.

Less babysitting means you hand it the job, go do something else, and come back to work that runs.

Very close, for half the price. On GDPval-AA, which tests real work from 44 jobs, Sonnet 5.5 scored 1844. Opus 5.5 scored 1846. On OSWorld 2.1, Sonnet scored 80.1% and Opus 81.8%. On CursorBench 4.0, 55.5% vs 57.8%.

Opus still comes out ahead on those. But a 2 point gap at twice the price is why Sonnet 5.5 is the one I'd run most of the day.

Sonnet 5.5 is the first Sonnet to beat Pokémon Red working only from screenshots. It looked at the screen and played.

Creator Kevin Ngo put it this way: "When Claude Opus 5.5 sets the architecture and general framework for a game, I would feel confident in letting Sonnet 5.5 implement it." That's a good way to split the work. Opus plans, Sonnet builds.

Anthropic gave Sonnet 5.5 a public company's earnings materials and a slide template. They asked for a 10-slide operating review. Two experts judged the first draft ready to send as is.

It also adds more polish to the screens and pages it builds, and it writes more clearly. Testers called it a better partner to work with. Every's Tyler Nishida summed it up in four words: "Claude Sonnet 5.5 cooks."

It came out today, so the builds are still rolling in. See what people made on day one on the interactive Sonnet 5.5 page.

Here's what this table says. Opus 5.5 wins every row except Terminal-Bench. Sonnet 5.5 wins that one, and it's the test closest to handing an AI a real job and walking away.

And here's what each model costs. Prices are per million tokens. In is what you send Claude. Out is what Claude writes back.

Anthropic says Opus 5.5 is clearly stronger at complex, open-ended work that needs steady judgment over a long stretch. Two more things to know about Sonnet 5.5:

On safety, Anthropic's containment tests found Sonnet 5.5 the least likely of any of their models to probe the limits of its test setup.

Here's how I'd split the work:

If you're on Pro, Max or Team, you also got one free usage reset. Use it any time before October 22.

Then give it one real job you'd normally sit and watch. See how far it gets alone. And if you want the visual version, watch the old and new Sonnet race side by side on the interactive page.