Claude Sonnet 5.5: same price, faster and cheaper per task

Anthropic released Claude Sonnet 5.5, and the pitch is efficiency rather than a new capability ceiling. The per-token price is unchanged from Sonnet 5, at $2 per million input tokens and $10 per million output, but Anthropic says the model runs more than 30% faster and finishes most tasks for up to 30% less because it needs fewer tokens to get there. The benchmark jumps are large on the agentic tests: Terminal-Bench 4.0 went from 10.3% to 70.6%, and CursorBench 4.0 from 34.1% to 55.5%. On coding it now lands close to Opus 5.5, which matters because Sonnet sits in the cheaper tier.

It also powers the free tier at claude.ai now. Simon Willison tested it through that free tier and had it generate a working WebGL 3D page, and he notes it is the first Sonnet model to beat Pokémon Red from screenshots alone. He also caught a rough edge: with thinking effort set to "max", the model burned through 128,000 thinking tokens, about $1.28, and then failed to produce any output. Setting it to "xhigh" gave a better answer for about six cents. The model is available now on AWS, Google Cloud, and Azure under the claude-sonnet-5-5 identifier.

Why it matters

If you already use Sonnet in production, this is close to a free upgrade: same price, faster, and fewer tokens per job, so it is worth testing whether it lets you move work down from Opus. But hold off on the top "max" thinking setting until that token-burn bug is fixed.