Claude Sonnet 5.5: A faster, more economical model

What's new

Anthropic launched Claude Sonnet 5.5 on September 28, 2026. The model is the second-largest in the Claude 5.5 family, just below the largest model, Opus 5.5. It is available in the Claude app and through the cloud platforms AWS, Google Cloud and Microsoft Azure. Zero data retention is also an option, which means data is not stored at Anthropic. Developers can find the model under the name "claude-sonnet-5-5".

Anthropic says the smaller sibling model, Claude Haiku 5.5, will arrive within the coming weeks. The model's knowledge cutoff is June 2026, meaning the point at which its knowledge ends.

Kiro, a developer tool, received experimental support for the model on October 2, 2026 for Pro, Pro+, Pro Max and Power subscribers. It runs in the US and in Frankfurt.

What's clever

Anthropic states that the model generates responses more than 30 percent faster than Sonnet 5. It also uses fewer tokens (chunks of text) per task. Users can also choose an effort level (low, medium or high) and so weigh cost against quality themselves.

In Anthropic's own measurements, the model scores 70.6 percent on Terminal-Bench 4.0, compared with 10.3 percent for Sonnet 5. In computer use (OSWorld 2.1), the result is 80.1 percent, reported as partial. On knowledge work (GDPval-AA), it scores 1844 versus 1846 for Opus 5.5, so nearly the same level as the more expensive model. Anthropic also highlights that it is the first Sonnet model to complete Pokémon Red using screenshots alone.

Customers quoted by Anthropic itself report gains: Box says the model was more accurate, 2.4 times faster and used 12 percent fewer tokens. Atlassian says Rovo Agents can run up to 30 percent faster, and Zendesk says tickets were handled 20 percent faster. Both the figures and the quotes come from Anthropic and have not been independently verified.

Cheaper or better?

The token price is unchanged from Sonnet 5: 2 dollars per million input tokens and 10 dollars per million output tokens. Cache reads cost 0.10 dollars and cache writes 2.50 dollars per million tokens. According to Anthropic, the savings instead come from the model using fewer tokens, so a task can cost up to 30 percent less.

In Kiro, the model costs 1.3 times as many credits as the previous Sonnet model, and the context window is 1 million tokens, with adaptive thinking turned on by default. Monthly prices for regular users are not stated in the sources. Comparisons with competitors outside Anthropic are not mentioned either.

What it's good for

According to the Taiwanese tech site Inside, the model is aimed at everyday tasks such as code bug fixes, document work and presentations, as well as design and visual layout. For complex, multistep tasks requiring long reasoning, the larger Opus 5.5 is recommended.

In practice, Sonnet 5.5 suits companies that want faster responses and a lower bill without switching to the most expensive model. Anthropic also highlights improved safety and cybersecurity safeguards on par with Opus 5.5.