Claude Sonnet 5.5

Claude Sonnet 5.5 is the second model in the Claude 5.5 family, offering a significant upgrade over its predecessor, Sonnet 5. This new model runs 30% faster and costs up to 30% less for most workloads. It is designed to complement Claude Opus 5.5, excelling at well-defined everyday tasks such as debugging code and creating polished documents, slides, and spreadsheets.

Claude Sonnet 5.5 also boasts strong design capabilities. With a score of 70.6% on Terminal-Bench 4.0, a coding evaluation, it surpasses Sonnet 5's performance by 10.3% and trails behind Opus 5.5's score by just two points on GDPval-AA, a test measuring real-world work across various occupations. Notably, Sonnet 5.5 performs exceptionally well on long-horizon work, including tasks like understanding and interacting with Pokémon Red from screenshots—it is the first Sonnet model to achieve this feat.

In terms of collaboration, Sonnet 5.5 writes more clearly and effectively than its predecessors, making it an ideal partner for teamwork. Moreover, it operates at faster speeds, generating outputs up to 30% quicker than Sonnet 5. Cost efficiency is another advantage of Sonnet 5.5, as it typically requires fewer tokens to complete tasks, resulting in up to 30% lower costs per task compared to its predecessor.

In terms of safety and alignment, Sonnet 5.5 maintains cybersecurity capabilities comparable to Opus 5, incorporating safeguards and fallbacks specifically developed for the most advanced models. It also includes biology safeguards, ensuring robust protection against high-risk requests while leaving routine software development and most life sciences work unaffected.

Though Sonnet 5.5 demonstrates impressive performance, particularly in coding and long-horizon tasks compared to Sonnet 5, it may not match Opus 5.5's capabilities in complex, open-ended tasks that require sustained judgment. However, at lower effort settings, Sonnet 5.5 often outperforms Sonnet 5 at a lower cost, making it a valuable complement to Opus 5.5 across various workloads.