Anthropic's Claude Sonnet 5.5 Is Out, Beats Opus 5.5 at Coding for Half the Price

What it claims

Sonnet 5.5 is Anthropic's mid-tier model, released as an upgrade to June's Sonnet 5. Anthropic says it runs more than 30% faster than Sonnet 5 and uses nearly a third fewer tokens per task, so it should cost less to run than Sonnet 5 at the same list price. Anthropic positions it for "well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets." Anthropic also says Opus 5.5 remains clearly stronger at complex work that needs sustained judgment.

Evidence

• Terminal-Bench 4.0 (share of agentic command-line tasks completed):
• Anthropic's self-reported scores: Sonnet 5.5 70.6%, Opus 5.5 66.4%, Sonnet 5 10.3%.
• Artificial Analysis's run: Sonnet 5.5 63.6%, Opus 5.5 59.6%, OpenAI GPT-6 Astra 59.1%.
• Artificial Analysis's tweet gives Sonnet 5.5 (max) about 64%, a 50-point gain over Sonnet 5 (max). It says the model also ranks near the top on Terminal-Bench-Science, but gives no number for it.
• FrontierCode: Anthropic says Sonnet 5.5 at High effort matches GPT-6 Sol at about one fifth of the cost per task. No score is given.
• GDPval-AA (Elo across 44 occupations): Sonnet 5.5 1844, Opus 5.5 1846, GPT-6 Sol 1487.
• Overall ranking: Artificial Analysis places Sonnet 5.5 second, behind Opus 5.5.

How to use it

• List price: $2 per million input tokens and $10 per million output tokens.
• Competing prices: OpenAI cut GPT-6 Sol to the same $2/$10 last week. GPT-5.6 Terra lists at $2/$12. Price alone no longer separates Sonnet 5.5 from GPT-6 Sol.
• Effort setting: this drives what you actually pay.
• Medium is the default in Anthropic's apps. Anthropic says it beats Sonnet 5's best coding score for less than a tenth of the cost.
• High is the setting Artificial Analysis calls the best value.
• Max cost $7.60 per task in Artificial Analysis's testing. That is about 50% more than Sonnet 5, and it reverses Anthropic's "up to 30% savings" claim.
• Haiku 5.5, aimed at high-volume, cost-sensitive workloads, is due in the coming weeks.

Limits

• Token use at max effort: about 193,000 tokens per task, the most Artificial Analysis has measured and roughly 60% more than Opus 5.5. Benchmark headline numbers are max-effort numbers. Budget for that if you run at max.
• Self-reported figures: Anthropic's table is its own. Its Terminal-Bench numbers run about 4–7 points above Artificial Analysis's for the same models.
• Pre-release build: Artificial Analysis tested a pre-release build that had a bug. Anthropic expects the bug changed little or slightly understated the scores.
• No comparison with GPT-5.6 Terra: Anthropic published no benchmarks against Terra, which is the closest-priced mid-tier rival.
• Missing figures: no FrontierCode scores and no Terminal-Bench-Science scores are given, only the relative claims above.