Anthropic Releases Sonnet 5.5 at $2 per Million Input Tokens — Half the Price of Opus 5.5, and It Beats It on Coding

Anthropic released Claude Sonnet 5.5 on Monday, the second model in its new Claude 5.5 family to reach customers in less than a week — and a telling one. The company says the mid-tier model does not push the frontier of its capabilities at all. It exists to do everyday work faster and cheaper than anything else Anthropic sells.

The pricing makes that positioning explicit. Sonnet 5.5 costs \$2 per million input tokens and \$10 per million output tokens — unchanged from Sonnet 5, but exactly half of Opus 5.5, which launched last week at \$4/\$20. Anthropic says the new model is also more than 30% faster than its predecessor and typically needs fewer tokens to finish the same task, cutting real-world cost per task by up to 30%.

The surprise is in the coding numbers. On Terminal-Bench 4.0, which scores whether an AI agent can complete complex professional tasks in a terminal, Anthropic's self-reported figures put Sonnet 5.5 at 70.6% — ahead of Opus 5.5's 66.4% and a dramatic jump from Sonnet 5's 10.3%. Independent tester Artificial Analysis recorded 63.6% for Sonnet 5.5 against 59.6% for Opus 5.5 and 59.1% for OpenAI's GPT-6 Astra. On GDPval-AA, which grades real-world professional work across dozens of occupations, the two Anthropic models finished in a statistical tie (1,844 vs 1,846).

"Sonnet is really for the cost-conscious customer where they might not need as much intelligence," Theo Chu, a research product manager at Anthropic, told CNBC. "It might be routine tasks that just need execution, but don't need that judgment that Opus can bring."

Because Sonnet 5.5 does not advance the frontier, most of its alignment testing focused on "a targeted set of risks that apply to models of any capability level." But its cybersecurity capabilities improved enough that Anthropic made it the first Sonnet to ship with the same fallbacks and cyber safeguards developed for its most capable models — risky requests fall back to Sonnet 5, and the model adds protections against distillation and reasoning-extraction attacks.

The model is available Monday across Anthropic's own products and on Amazon Web Services, Google Cloud and Microsoft Azure. Anthropic says enterprise customers account for roughly 80% of its business, including Salesforce, Databricks, Goldman Sachs and Novo Nordisk.

The launch lands in the middle of a fast-compressing mid-tier price war: OpenAI cut GPT-6 Sol to the same \$2/\$10 last week. There is a caveat buried in the benchmarks — Artificial Analysis found Sonnet 5.5 at maximum effort wrote about 193,000 tokens per task, the most it has ever measured, costing roughly \$7.60 per task. The efficiency story only holds when the effort dial stays low, which is precisely how Anthropic intends most customers to use it. The cheapest member of the family, Haiku 5.5, is due "in the coming weeks."