Anthropic releases Claude Sonnet 5.5, a faster and cheaper follow-up to Opus 5.5

Anthropic has unveiled Claude Sonnet 5.5, the faster and more cost-effective sibling to its Opus 5.5 model, less than a week after launching the latter. According to the company's announcement, Sonnet 5.5 operates more than 30% quicker than its predecessor and comes at a discounted price, being up to 30% cheaper for most tasks. Described as a more affordable counterpart to Opus 5.5, Sonnet 5.5 shines in well-defined everyday tasks like fixing bugs and creating documents, slides, and spreadsheets. In the near future, Claude Haiku 5.5 is expected to join the lineup.

Benchmark results show Sonnet 5.5 delivering significant improvements over its predecessor, though slightly trailing behind Opus 5.5 at various performance levels. On GDPval-AA v2.1, which evaluates real-world work across 44 occupations, Sonnet 5.5 scored 1,844, just two points below Opus 5.5 and about 400 points higher than Sonnet 5. The model also managed to surpass Pokémon Red, relying solely on screenshots.

Anthropic emphasized that benchmarks only scratch the surface of a model's capabilities and noted that Opus 5.5 still outperforms Sonnet 5.5 in complex, open-ended tasks during internal and external testing. Early users have reported efficiency gains, with Slack's principal engineer, Curtis Allen, reporting that Sonnet 5.5 outperformed Sonnet 5 on most of Slack's offline Slackbot evaluations, completing tasks in fewer steps and generating about 14% fewer output tokens.

Pricing remains consistent with Sonnet 5, with input and output rates set at $2 and $10 per million tokens, respectively, and a $0.20 charge per million tokens for cache reads. The lower per-task cost is attributed to the model's ability to complete tasks with fewer tokens. Sonnet 5.5 is now available on all of Anthropic's platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure, and can be accessed via the API as claude-sonnet-5-5 without any data retention.

Developers who run Sonnet with thinking off must switch to a new between_tools setting before migrating.

Anthropic confirmed that Sonnet 5.5's cybersecurity capabilities are comparable to those of Opus 5, making it the first Sonnet model to include cyber safeguards similar to those found in Opus 5.5. Higher-risk cybersecurity requests will be automatically redirected to Sonnet 5, while routine bug-fixing remains unaffected. The model also retains biology safeguards from Sonnet 5, with broader access granted upon application to the Life Sciences Verification Program.

Additionally, Sonnet 5.5 introduces classifiers designed to thwart distillation attacks, which enable malicious actors to extract a model's capabilities at scale. In Anthropic's automated behavioral audit of approximately 1,850 scenarios, Sonnet 5.5 matched or improved upon Sonnet 5 on most alignment measures, although Opus 5.5 still achieved slightly better results overall. The company acknowledged that no set of evaluations can account for every potential failure.