Anthropic released Claude Sonnet 5.5 on Monday, framing its mid-tier model as the practical pick for routine work. The pitch rests on two numbers: answers come back more than 30 percent faster than Sonnet 5, and the model needs far fewer tokens to finish the same job.
List pricing does not move. Input stays at $2 per million tokens, output at $10, and cache reads at $0.20. The saving comes from token efficiency, which Anthropic puts at roughly 30 percent per completed task.
Benchmarks tell a sharper story. Sonnet 5.5 scored 70.6 percent on Terminal-Bench 4.0, an agentic coding test, against 10.3 percent for the model it replaces. It also beat the larger Opus 5.5 on agentic coding while landing about two points behind it on GDPval-AA, a measure of real occupational work. Anthropic says it is the first Sonnet to finish Pokemon Red.
The model inherits the foundation behind Opus 5.5, which makes it the first Sonnet with meaningful cyber capability. That triggers the same safeguards and fallbacks Anthropic applies to Fable and Mythos-class systems, and they can fire on cyber, biology or other sensitive prompts.
Anthropic also built invisible text watermarking into the release, part of a push to satisfy global content rules without changing how the writing reads. Early customers report faster handling of routine work. Zendesk’s Abhinay Kathuria said its support queue moved about 20 percent quicker and made fewer wrong calls.
Haiku 5.5, the smallest tier in the family, follows in the coming weeks.
