Anthropic releases Sonnet 5.5, beating Opus 5.5 at coding for half the price

Anthropic released Claude Sonnet 5.5 on Monday. The model upgrades Sonnet 5, which launched in June. Anthropic says Sonnet 5.5 runs more than 30% faster than Sonnet 5. Anthropic set its price at $2 per million input tokens and $10 per million output tokens. Those prices are half of Opus 5.5's prices. Sonnet 5.5 uses nearly a third fewer tokens per task than Sonnet 5, which makes it cheaper to run. On Terminal-Bench 4.0, Sonnet 5.5 scored 70.6%. Opus 5.5 scored 66.4% on the test. Sonnet 5 scored 10.3%. Terminal-Bench 4.0 measures whether an AI agent can complete complex professional tasks by typing commands independently. Artificial Analysis, an independent testing firm, scored Sonnet 5.5 at 63.6% on its version of the test. Artificial Analysis scored Opus 5.5 at 59.6%. It scored OpenAI's GPT-6 Astra at 59.1%. Anthropic says Sonnet 5.5 at High effort matches GPT-6 Sol on FrontierCode for about a fifth of the cost per task. On GDPval-AA, Sonnet 5.5 scored 1844. The test measures professional work across 44 occupations using Elo rankings. Opus 5.5 scored 1846, which the article describes as effectively a tie with Sonnet 5.5. GPT-6 Sol scored 1487. Artificial Analysis ranks Sonnet 5.5 second behind Opus 5.5. Artificial Analysis says Sonnet 5.5 used more tokens per task than any model it tested. At maximum effort, Sonnet 5.5 generated about 193,000 tokens per test task, the most Artificial Analysis has measured. Each task cost $7.60 at that setting. That was about 50% more than Sonnet 5's cost per task and cuts against Anthropic's claim of up to 30% savings. Anthropic says Sonnet 5.5 at Medium effort, the default in its apps, beats Sonnet 5's best coding score for less than a tenth of the cost. Artificial Analysis says High effort offers the best value. OpenAI cut GPT-6 Sol's price to $2 per million input tokens and $10 per million output tokens last week. GPT-5.6 Terra lists at $2 per million input tokens and $12 per million output tokens. Anthropic published no benchmarks for Terra. Anthropic's table is self-reported. Artificial Analysis tested a pre-release build with a bug. Anthropic expects the bug changed little or slightly understated its scores. Anthropic says Opus 5.5 remains clearly stronger at complex work requiring sustained judgment. Claude Haiku 5.5 is built for high-volume, cost-sensitive applications and is due in the coming weeks.