Anthropic Releases Claude Sonnet 5.5 With Lower Costs and Stronger Coding Performance

Summary

Anthropic has released Claude Sonnet 5.5, positioning it as a faster and less expensive alternative to the company’s Opus 5.5 model. The article says Sonnet 5.5 is more than 30% faster than Sonnet 5 and can reduce the actual cost of a task by up to 30% because it uses fewer tokens, although its listed price remains $2 per million input tokens and $10 per million output tokens, with cached input priced at $0.20. The model supports five adjustable effort levels, allowing users to trade response speed and token use for deeper checking on difficult tasks. Reported evaluations place it ahead of Opus 5.5 on Terminal-Bench 4.0, at 70.6% versus 66.4%, while it also posted stronger or near-parity results on coding, office-work and long-context knowledge-work tests. The article cites scores including 55.5% on CursorBench, 1,844 on GDPval-AA, 1,811 on AA-Briefcase and 80.1% on OSWorld 2.1. It also describes improvements in chart interpretation, screenshot-based computer control, tool-call batching, interface design and presentation generation. In an internal example, the model produced a 10-page operational-review presentation that two human experts judged ready to send. Demonstrations included building a 3D San Francisco scene, interactive models and a small Game Boy-style game, although one separate Lego-world challenge reportedly failed. The model’s visual performance was also illustrated by completing Pokémon Red from screenshots alone. The article warns that maximum effort can erase the cost advantage: one test measured an average of 193,000 output tokens per task and a $7.60 average cost, above Opus 5.5’s reported $5.98. Maximum effort also performed worse than a lower setting on one coding benchmark. Anthropic says Sonnet 5.5 adds a safety classifier intended to prevent reasoning extraction and introduces Opus-level cybersecurity safeguards, including fallback to Sonnet 5 for high-risk requests.