Claude Sonnet 5.5 runs 30% faster at the same token price

Anthropic released Claude Sonnet 5.5, the second model in the 5.5 family, as a faster and cheaper complement to Claude Opus 5.5 rather than a replacement for it. Claude Haiku 5.5, aimed at high volume and cost sensitive use, is promised “in the coming weeks.”

The headline claim is efficiency, not a capability jump. Sonnet 5.5 generates output more than 30% faster than Sonnet 5, which Anthropic calls its fastest Sonnet to date, and the list price is unchanged at $2 per million input tokens, $10 per million output, $0.20 per million cache reads and $2.50 per million cache writes. Opus 5.5 costs $4, $20, $0.20 and $5. The savings come from needing fewer tokens, not a lower rate: Anthropic says it costs up to 30% less per task in its own testing.

The numbers that moved

• Terminal-Bench 4.0: 70.6%, against 10.3% for Sonnet 5 and 66.4% for Opus 5.5 at Xhigh effort. On this one it edges the larger model.
• CursorBench 4.0: 55.5% against 34.1%, with Opus 5.5 at 57.8%.
• GDPval-AA v2.1: 1844 against 1449, landing two points under Opus 5.5’s 1846. AA-Briefcase v1.1: 1811 against 1359.
• Humanity’s Last Exam: 64.5% with tools. OSWorld 2.1: 80.1% partial. Chartography: 61.6% with no tools, against 15.6% for Sonnet 5.
• First Sonnet model to beat Pokémon Red working only from screenshots.

The more interesting claim is about cost per task rather than score. At Low or Medium effort, Sonnet 5.5 beats Sonnet 5’s best score for roughly a tenth of the cost per task on Terminal-Bench and CursorBench, and about a ninth on AA-Briefcase. On FrontierCode at High effort it matches GPT-6 Sol’s best score for about a fifth of the cost. Early testers attributed the gain to batching tool calls together, which cuts both steps and tokens.

What to watch

Because Sonnet 5.5’s cyber capabilities are comparable to Opus 5’s, it is the first Sonnet to ship with Opus grade cyber safeguards, and higher risk cybersecurity requests visibly fall back to Sonnet 5. It is also the first Sonnet to launch with classifiers against reasoning extraction from distillation attacks, and it expands preserved thinking so a session’s reasoning cannot be moved to another account.

Anthropic also lists four caveats against its own numbers, including a pre-release structured output bug that degraded the Artificial Analysis GDPval-AA and AA-Briefcase runs and a Max effort timeout penalty on FrontierCode. The model is live on all platforms including AWS, Google Cloud and Azure with zero data retention, as claude-sonnet-5-5. Anyone running Sonnet with thinking off has to move to the new between_tools setting first.

(Original source: Anthropic)