Anthropic shipped Claude Sonnet 5.5 on September 28, six days after Opus 5.5, holding list prices flat at $2 per million input tokens and $10 per million output tokens while claiming the model completes the same work with up to 30% less cost per task. SiliconAngle's writeup is direct about the mechanism: token efficiency is the entire source of the savings. Anthropic itself puts it as follows: the model "typically needs far fewer tokens to do the same work."

Sticker prices didn't move. The bill does, if the pipes cooperate.

The benchmark picture is unusually lopsided. Sonnet 5.5 posts 70.6% on Terminal-Bench 4.0 against Sonnet 5's 10.3%, a generational jump that in any prior release cycle would be positioned as flagship territory. It also clears Opus 5.5's 66.4% at its highest effort setting, per The Next Web, despite Opus costing twice as much per token. On GDPval-AA v2.1, Sonnet 5.5 lands at 1,844, essentially tied with Opus 5.5's 1,846 and well above Sonnet 5's 1,449. The mid-tier has effectively absorbed what used to be the top-tier price-performance frontier.

Anthropic's launch materials are stacked with customer numbers that all point the same direction. Slack principal engineer Curtis Allen reports roughly 14% fewer output tokens on offline Slackbot evaluations with no prompt changes. Box reports results 2.4 times faster with 12% fewer total tokens. Balyasny Asset Management, running an evaluation across 2,441 finance tasks, reports about 121,000 tokens per answer versus Sonnet 5's 497,000. Base44's Gabriel Grinberg reports 3.6 iterations per build against Opus 5's 7.7. Zendesk's Abhinay Kathuria clocks a 20% faster ticket-processing improvement over Claude models already in production.

These are vendor-curated figures, and they should be read that way. But the shape is consistent: same job, fewer tokens, fewer round trips.

The pricing context matters. Sonnet 5's $2/$10 rates were introductory, scheduled to rise to $3/$15 after August 31. Anthropic is now holding the promotional floor and layering an efficiency claim on top, which is a more disciplined move than raising rates on a refresh. There's a rougher edge: sending Sonnet 5.5 a disabled-thinking setting returns a 400 error, meaning integrations built for prior models need config changes before they'll run.

Haiku 5.5 is next, per TechCrunch's read of Anthropic's roadmap. Until it arrives, Sonnet 5.5 is the model that quietly redraws the budget line for anyone running production Claude workloads.

Sources

https://www.anthropic.com/claude-sonnet-5-5
https://techcrunch.com/2026/09/28/anthropic-releases-sonnet-5-5-which-it-calls-a-significantly-cheaper-faster-work-partner/
https://siliconangle.com/2026/09/28/anthropic-debuts-claude-sonnet-5-5-running-30-faster-than-the-previous-generation-ai-model/
https://thenextweb.com/news/sonnet-5-5-cyber-distillation
https://www.unite.ai/anthropic-releases-claude-sonnet-5-5-at-unchanged-sonnet-5-pricing/