Sonnet 5.5 edges past Opus 5.5 on Terminal-Bench 4.0

Anthropic released Claude Sonnet 5.5 on September 28 with a Terminal-Bench 4.0 score of 70.6%, ahead of Opus 5.5’s 66.4%. That comparison puts Sonnet at Max effort against Opus at Xhigh. At every other matching effort level, Sonnet trails Opus on that benchmark, Anthropic’s announcement and system card show.

Sonnet 5 scored 10.3% on the same agentic coding test, and Sonnet 5.5 keeps its price of $2 per million input tokens and $10 per million output. Anthropic says the new model costs up to 30% less per task because it needs fewer tokens for the same work. It also generates output 30%+ faster, which makes it the fastest Sonnet so far. Anthropic pitches it at well-scoped everyday work: bug fixing, documents, slides and spreadsheets. In Anthropic’s own words, Opus 5.5 “remains clearly stronger at complex, open-ended work requiring sustained judgment.”

Same price, fewer tokens

At launch Sonnet 5.5 charged $2.50 per million tokens for cache writes and $0.20 for cache reads, against $5 and $0.20 on Opus 5.5, which also charges $4 input and $20 output. Later, on October 7, the day Haiku 5.5 arrived, Anthropic halved Sonnet 5.5’s cache reads to $0.10 per million and said this makes the model “around 20% cheaper to run on most long-running work.” So every row of Sonnet’s price list is now exactly half of Opus’s.

Since the rates did not change, most of the per-task saving comes from how the model works. Early testers saw it batch tool calls together more often than Sonnet 5, so a task takes fewer steps. Customers put numbers on it: Balyasny Asset Management measured about 121K tokens per answer against 497K on Sonnet 5, and Base44 reported 3.6 iterations per app build where Opus 5 took 7.7. Slack ran its offline Slackbot evals without touching the prompts.

“Without changing any of our prompts, Claude Sonnet 5.5 did better than Sonnet 5 on almost all of our offline Slackbot evals, in fewer steps and with about 14% fewer output tokens.” (Curtis Allen, principal engineer, Slack)

Epic Games COO Daniel Vogel said the model handled tens of thousands of lines of gameplay system code, “kept responses snappy, handled multi-hour tasks, and delivered with less prescriptive prompting.”

Which effort level to run

Effort has five levels: low, medium, high, xhigh and max. Claude Code and the Claude apps default to Medium, the Claude Platform to High. Anthropic says that on several benchmarks Sonnet 5.5 at Low or Medium beats Sonnet 5’s best score for about a tenth of the cost per task. On Terminal-Bench, Medium on its own clears Sonnet 5’s best for less than a tenth of the cost. On FrontierCode at High, it scores 10 points above Sonnet 5 at the same setting, at about one fifteenth of the cost.

The top setting is not always the best one. On FrontierCode, Sonnet 5.5 scores 46.2% at Max and 52.1% at Xhigh. MarkTechPost says that at Max the model more often ran multi-agent code review, which sometimes led to timeouts or out-of-scope edits that the benchmark penalizes.

On GDPval-AA, which covers tasks from 44 occupations, Sonnet 5.5 scores 1844 against Opus 5.5’s 1846 and Sonnet 5’s 1449. On CursorBench, built from real Cursor sessions, it gets 55.5% against Opus’s 57.8%. In one internal test it turned a public company’s earnings materials and a slide template into a 10-slide operating review, and two experts judged the first draft ready to send. Anthropic’s own guidance is that Sonnet complements Opus best at lower effort settings.

Before you switch

The model ID is claude-sonnet-5-5. It is available on all platforms, including AWS, Google Cloud and Azure, and with zero data retention. It is the first Sonnet to ship with cyber safeguards like those on Opus 5.5. Routine bug fixing is unaffected, but higher-risk security tasks visibly fall back to Sonnet 5. Cyberdefenders will soon be able to apply to an expanded Cyber Verification Program for tiered access. The biology safeguards are the same as on Sonnet 5, and some microbiology and virology requests may be flagged in error. Organizations can apply to the Life Sciences Verification Program for safeguards that cover the full range of biology work.

To block distillation (attackers using fake accounts to pull out a model’s reasoning), Sonnet 5.5 keeps its thinking tied to the organization that generated it. If you switch accounts mid-session in Claude Code, Claude rereads the session and generates new thinking. Separately, Pro, Max and Team users can still apply the reset Anthropic granted them at any time before October 22.

If you run Sonnet with thinking off, switch to the new between_tools setting, which keeps up-front thinking off, before moving to Sonnet 5.5. The migration guide covers the details.