Anthropic released Claude Sonnet 5.5 on September 28 and priced it at $2 per million input tokens and $10 per million output tokens. The bigger claim is about how many tokens a task consumes, which is what a buying committee actually pays for.

According to Anthropic’s announcement, Sonnet 5.5 runs 30% faster than Sonnet 5 and costs up to an estimated 30% less to run on typical token-billed workloads. The per-token rate matches the prior generation, so the saving comes from the model finishing the same work in fewer tokens. The model is available on the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Foundry, and Anthropic says a Haiku model is due in the coming weeks.

Why it matters to the buying committee

Model price lists still quote a rate per token, while budgets get spent on tokens per completed task. The customer figures Anthropic publishes show the gap. Slackbot’s engineering team reports about 14% fewer output tokens than Sonnet 5 on its offline evals. One finance-focused customer reports roughly 121,000 tokens per answer on a 2,441-task suite, where Sonnet 5 used 497,000. These are vendor-selected testimonials on the vendor’s own page, so treat them as a starting hypothesis for your own workload rather than a benchmark.

The original angle

If the same rate card produces a different bill each generation, contracts written around committed spend per token are harder to forecast than contracts written around committed spend per outcome. Procurement teams renewing model commitments this quarter should ask vendors for tokens-per-task data on a sample of their own workflows, and should re-run that measurement at every model upgrade. A 30% drop in token use is a 30% cut in the committed volume you needed to buy.

Source: Anthropic, Claude Sonnet announcement page