Anthropic's Claude Sonnet 5.5: 30% faster, clears Pokémon Red

Claude Sonnet 5.5 arrives as part of Anthropic's Claude 5.5 family, positioned as the faster and more efficient option compared with the Claude Sonnet 5 it replaces. The pricing is unchanged at $2 per million input tokens and $10 per million output tokens, but Anthropic attributes the cost improvement to using fewer tokens for the same task rather than to a price reduction.

The performance jump on Terminal-Bench 4.0, from 10.3% to 70.6%, is unusually large for a model refresh, and the article says the model also approached Opus 5.5 on GDPval-AA. Anthropic is presenting the model's long-horizon task ability and image recognition as strengths, pointing to it being the first Sonnet model to clear Pokémon Red using only screenshots. That claim speaks less to typical business workloads and more to the model's ability to act over extended sessions with visual input.

The security additions are worth noting. Anthropic introduced safeguards for high-risk tasks as its cybersecurity capabilities improved, and added safety classifiers aimed at distillation attacks that try to improperly extract a model's capabilities. The article says these restrictions are applied in a way that does not affect general software development or biology research. Whether the efficiency claims translate into real savings will likely hinge on how much customers' own task patterns resemble the token-saving behavior Anthropic describes.