Anthropic Launches Claude Sonnet 5.5: Same Price, 30% Faster
Anthropic says the new model generates answers more than 30% faster than its predecessor. Per-token pricing is unchanged, but by the company's own measurements the cost of getting the same job done falls by up to 30%.
Anthropic released Claude Sonnet 5.5 on September 28 (local time), the second model in the Claude 5.5 family. It is aimed at well-scoped everyday work: fixing bugs in code, and drafting documents and slide decks. Anthropic says it generates responses more than 30% faster than Claude Sonnet 5, and that in the company’s own testing, the cost of completing a single task fell by as much as 30%.
Same sticker price, lower bills
For developers wiring the model into their own products, the per-token price is unchanged from the previous model: $2 per million input tokens and $10 per million output tokens. Reading back cached input costs $0.20 per million tokens. Tokens are the small chunks of text an AI uses to process language. Anthropic says the real-world cost per task drops because the model needs fewer tokens to finish the same job.
Users can trade off cost and speed by dialing how much effort the model spends reviewing its own answers. Set that review effort low and it responds faster on routine work; set it high and it deliberates longer. In Claude Code and the app, the default is the middle setting, Medium. On Claude Platform, the developer-facing product, the default is High. That means an identical request can take different amounts of time and money depending on where you run it.
A coding score of 70.6%, and what changed
Coding tests showed a large gap between Sonnet 5.5 and its predecessor. On Terminal-Bench 4.0, which measures how well an AI handles coding work on its own, Sonnet 5.5 scored 70.6%. Sonnet 5 scored 10.3%. Anthropic also passed along feedback from early testers, who reported that the model’s ability to juggle several tools at once cut both the number of steps and the cost of a task.
Image: anthropic.com
The company also highlighted gains in writing documents and preparing materials. On GDPval-AA, which evaluates real tasks drawn from 44 occupations across nine major industries, Sonnet 5.5 came in 2 points below Opus 5.5. In an internal test, Anthropic handed the model a public company’s quarterly earnings materials, an earnings call transcript and a slide template, and asked it to produce a 10-slide review deck. Two experts judged the first draft good enough to send as-is, without edits.
Anthropic says that when the review effort is pushed to its maximum, Max, Sonnet 5.5 performed comparably to Opus 5.5 on some benchmarks. Still, the company says Opus 5.5 remains stronger on complex work that requires sustained judgment and on open-ended tasks with no fixed answer. It added that raising Sonnet 5.5’s review effort can push its per-task cost close to that of Opus 5.5.
Where you can use it
Anthropic says Sonnet 5.5 is available on every platform it supports, including Amazon Web Services (AWS), Google Cloud and Microsoft Azure. Developers can also use it on Claude Platform. Developers who had been running an earlier Sonnet with answer review turned off will need to change that setting when migrating to the new model.
Image: anthropic.com
New safeguards apply to security-related requests. Anthropic says the model will continue to handle ordinary software development and debugging, but will route higher-risk cybersecurity work to Sonnet 5 instead. For biology, it uses the same safeguards as Sonnet 5. The company acknowledges that this can occasionally cause legitimate microbiology and virology requests to be misclassified as risky.
Anthropic also said Claude Haiku 5.5, aimed at high-volume and cost-sensitive uses, will join the family within the next few weeks. The point of this release is straightforward: finish repetitive work faster, and spend less on each task.