Reviews · SEPTEMBER 30, 2026
Claude Sonnet 5.5 Lands at Sonnet 5 Prices, Jumps Terminal-Bench 4.0 to 70.6%
Anthropic's mid-tier model ships 30%+ faster and up to 30% cheaper per task at unchanged $2/$10 list pricing, with five recalibrated effort levels and five breaking API changes to clear before migrating.
Anthropic shipped Claude Sonnet 5.5 on September 28, 2026, holding list pricing at $2 per million input tokens and $10 per million output tokens while claiming 30%+ faster generation and up to 30% lower cost per task versus Sonnet 5. The headline number is a Terminal-Bench 4.0 score of 70.6%, up from Sonnet 5's 10.3%. That isn't a generational bump. It's the tier catching up to what the prior Opus release had already made table stakes for agentic coding.
The pricing hold is the strategic tell. Anthropic is protecting the price line and letting the efficiency gains translate directly into cheaper unit economics for anyone already running Sonnet 5 in production. At Low or Medium effort, Sonnet 5.5 hits roughly a tenth the cost per task of Sonnet 5's best score on Terminal-Bench 4.0, CursorBench 4.0, and AA-Briefcase v1.1. For small teams whose sales, content, and agent pipelines are already Sonnet-shaped, that's the entire migration case.
The five newly exposed effort levels, low, medium, high, xhigh, max, are where the routing decision lives. On FrontierCode 1.1 (main), Sonnet 5 posts 42.4%, Sonnet 5.5 climbs to 52.1% at Xhigh and 46.2% at Max, and Opus 5.5 tops the field at 54.4%. CursorBench 4.0 shows the same shape: Sonnet 5.5 at 55.5%, Opus 5.5 at 57.8%. Since Opus 5.5 lists at $4/$20, exactly 2x the Sonnet 5.5 rate, Max-effort Sonnet runs can quietly cost more per completed task than Opus. Route accordingly.
The launch partners tell a consistent efficiency story. Slack's Curtis Allen reports ~14% fewer output tokens in offline Slackbot evaluations. Box is 2.4x faster with 12% fewer total tokens. Zendesk's Abhinay Kathuria cites 20% faster ticket processing against incumbent Claude models. Atlassian's Rovo Agents run up to 30% faster than on Sonnet 5. Balyasny Asset Management, across 2,441 finance tasks, saw per-answer token spend fall from 497,000 to roughly 121,000, a ~76% reduction. Base44's Gabriel Grinberg reports 3.6 iterations per app build versus 7.7 on Opus 5, across 118 builds.
Then the fine print. There are five breaking API changes to clear before migrating. Disabled-thinking configurations now return a 400. The minimum cacheable prompt drops from 1,024 tokens to 512. Five refusal stop-reason categories now surface where fewer did before. Anyone whose retry logic or observability layer wasn't built for those signals will notice.
Distribution is already broad: Claude Code, the Claude apps, the Claude Platform, AWS, Google Cloud, Azure, and GitHub Copilot across Pro, Pro+, Max, Business, and Enterprise tiers via VS Code. Haiku 5.5 is next, which means this benchmark leaderboard has a short shelf life. The migration window is now.