This article contains affiliate links. As an Amazon Associate I earn from qualifying purchases.
Anthropic released Claude Sonnet 5.5 on September 28, 2026. It’s the second model in the Claude 5.5 family, following Opus 5.5, which came out on September 22. Pricing is unchanged from Sonnet 5, output is over 30% faster, and the cost per task is up to 30% lower, according to the announcement.
Full disclosure up front: I use Claude for research and proofreading when writing articles for this blog. This is a new version of a tool I use every day, so read what follows with that in mind.
Faster and Cheaper at the Same Price
API pricing is $2 per million input tokens, $10 per million output tokens, and $0.20 per million for cache reads, exactly the same as Sonnet 5. Next to Opus 5.5, that’s precisely half on both input and output.
How can it be “up to 30% cheaper” when the unit price hasn’t moved? Because it needs fewer tokens to finish the same job. One company that tested it early reported that across 2,441 finance tasks, Sonnet 5 used about 497,000 tokens per answer while Sonnet 5.5 got by with about 121,000. Another wrote that in its coding evals, tool calls dropped by a third and shell runs were roughly halved. Fewer steps means less waiting and lower bills. It’s as simple as that.
On speed, Anthropic says output generation is over 30% faster than Sonnet 5, making it the fastest Sonnet to date.
Coding Scores Jumped by an Order of Magnitude
The most striking gains are in coding. On Terminal-Bench 4.0, which tests multi-step work in a command-line environment, Sonnet 5.5 scored 70.6%. Sonnet 5 scored 10.3%. That gap is hard to believe for two models in the same line.
On CursorBench 4.0, built from tasks taken from real Cursor sessions, it hit 55.5%, up sharply from Sonnet 5’s 34.1% and within a little over two points of Opus 5.5’s 57.8%. On FrontierCode 1.1, which checks whether code changes can be merged as is, it scored 10 points higher than Sonnet 5 at the High setting, at about one fifteenth of the cost per task.
Beyond coding, it was nearly tied with Opus 5.5 on GDPval-AA, which measures real-world tasks across 44 occupations (1844 vs. 1846), and came close to Opus 5.5 on OSWorld 2.1 (computer use) and chart reading. It’s also the first Sonnet model to beat Pokémon Red working only from screenshots, which Anthropic cites as an example of improved image understanding and sustained long-horizon work.
A quieter improvement is that it writes more clearly. The same was said of Opus 5.5, and early testers described it as a better collaborator than Sonnet 5. It’s good at building documents that follow slide templates: in one internal test, it was given a company’s earnings materials and call transcripts and asked for a 10-slide deck, and two experts judged the first draft ready to send as is.
It Beats the Higher-Tier Opus 5.5 on Terminal-Bench
Here’s the number that stands out. On Terminal-Bench 4.0, Opus 5.5 scored 66.4%, which is lower than Sonnet 5.5’s 70.6%. The mid-tier model beat the top-tier model, at half the price.
Yet Anthropic itself warns against taking that number at face value. The announcement states that benchmark scores capture only one facet of a model’s capabilities, and that in both internal testing and external testers’ evaluations, Opus 5.5 remains clearly stronger at complex, open-ended work requiring sustained judgment. In a launch post for a new product, the company went out of its way to note where its higher-tier model still wins.
That line runs straight through the guidance on when to use which. Opus 5.5 is for complex work requiring careful judgment; Sonnet 5.5 is for well-scoped everyday tasks, bug fixes, and producing documents, slides, and spreadsheets. AWS’s write-up put it as handing Sonnet 5.5 the work where the approach is already clear and what’s left is to execute it quickly. One tester described the same pairing: let Opus 5.5 set a game’s architecture and framework, then have Sonnet 5.5 implement it.
The effort settings point the same way. At Low or Medium, Sonnet 5.5 beats Sonnet 5’s best score on several benchmarks for about a tenth of the cost per task. At higher settings, it performs comparably to Opus 5.5 at a comparable cost. In other words, the lower you run it, the more the price gap over Opus works in your favor, which is Anthropic’s own framing.
One more thing arrived with the performance jump. Because its cybersecurity capabilities are now on par with Opus 5, Sonnet 5.5 is the first Sonnet to ship with the same kind of safeguards used on Opus 5.5. Routine bug fixing and development are unaffected, but higher-risk security requests visibly fall back to Sonnet 5. It’s also the first Sonnet with classifiers to block distillation attacks, in which reasoning is extracted through thousands of accounts to build a separate model. That a mid-tier model now needs this level of protection says something about how far it has come.
Availability and Migration Notes
Sonnet 5.5 is available in the Claude web, iOS, and Android apps, and via the API under the model name “claude-sonnet-5-5.” It launched the same day on Amazon Bedrock, Google Cloud, and Microsoft Foundry. The context window is 1M tokens, max output is 128K tokens, and the knowledge cutoff is June 2026.
If you’ve been running Sonnet with thinking turned off through the API, you’ll need to switch to the new “between_tools” setting before moving to Sonnet 5.5. There’s also a change in response shape, with text between tool calls now returned in thinking blocks, so anyone with an existing integration should check the migration guide.
The last member of the family, Haiku 5.5 for high-volume work, is expected “in the coming weeks.”
Personally, what stuck with me wasn’t the Terminal-Bench upset but the developer spelling out the limits of its own benchmarks. As the numbers converge, the choice comes down to the kind of work. Sonnet 5.5 for getting decided things done fast, Opus 5.5 for when the deciding is part of the job. That’s Anthropic’s own summary.
Sources: Introducing Claude Sonnet 5.5 (Anthropic), Claude Sonnet 5.5 (Claude Platform Docs), Introducing Claude Sonnet 5.5 on AWS (AWS Machine Learning Blog)
Related products
You can find books and products related to this article on Amazon.
If you buy through the links above, this blog may earn a commission.
コメント