Claude Sonnet 5.5 Lands at the Same Price, and Says It Costs 30% Less to Use

Anthropic has released Claude Sonnet 5.5, the second model in its Claude 5.5 family, and the pricing is the part worth reading twice. Per-token rates are unchanged from Sonnet 5 — $2 per million input tokens, $10 per million output, $0.20 per million for cache reads — yet Anthropic’s announcement claims the model costs up to 30% less per task. The saving comes from finishing work in fewer tokens rather than from a discount.

It also generates output more than 30% faster than Sonnet 5, which Anthropic describe as making it their fastest Sonnet to date.

Where it sits in the line-up is clear enough: Sonnet 5.5 is the lower-cost complement to Claude Opus 5.5, aimed at well-scoped everyday work — fixing bugs, building documents, slides and spreadsheets — while Opus stays the model for tasks needing sustained judgement. Anthropic say Claude Haiku 5.5 will join the family in the coming weeks.

The coding numbers

The headline result is Terminal-Bench 4.0, an agentic coding evaluation, where Sonnet 5.5 scores 70.6% against Sonnet 5’s 10.3%. That is not an incremental gain, and it is the clearest evidence that the generation changed something structural rather than tuning the edges. Opus 5.5 records 66.4% on the same test, so the cheaper model is ahead of the expensive one here.

Elsewhere the ordering is more conventional. CursorBench 4.0 puts Sonnet 5.5 at 55.5%, up from 34.1%, with Opus 5.5 at 57.8%. On FrontierCode 1.1 it reaches 46.2% at Max effort against Sonnet 5’s 42.4%, while Opus 5.5 takes 54.4% and GPT-6 Sol 49.3%, rising to 52.1% at xhigh.

Knowledge work, where it nearly matches Opus

On GDPval-AA v2.1, a test of real-world work across a range of occupations, Sonnet 5.5 scores 1844 against Sonnet 5’s 1449 — and Opus 5.5’s 1846. Two points separate them. AA-Briefcase v1.1 tells the same story at 1811 against 1359, with Opus 5.5 on 1822 and GPT-6 Sol on 1483.

Other measures, all with tools enabled where noted: Humanity’s Last Exam at 64.5% with tools, up from 54.9%, with Opus 5.5 on 67.7%. OSWorld 2.1 at 80.1% partial against 57.0%, where Opus 5.5 records 81.8%. Chartography at 61.6% without tools, up sharply from 15.6%, with Opus 5.5 on 64.4%.

Anthropic add a caveat of their own to all of it: benchmark scores capture one facet of a model, and in their testing and their external testers’ testing, Opus 5.5 remains clearly stronger on complex, open-ended work.

Safety, and a first for the Sonnet line

Sonnet 5.5 is the first Sonnet model to launch with cyber safeguards of the kind Anthropic built for their most capable models, because its cybersecurity ability is comparable to Opus 5’s. Its biology safeguards match Sonnet 5’s. Anthropic note both target a narrow band of high-risk requests, and that routine software work and most life-sciences work are unaffected. On their automated behavioural audit the model improves on or matches Sonnet 5 across most alignment measures.

One more detail that will please anyone who has watched the benchmark: it is the first Sonnet to finish Pokémon Red working only from screenshots.