Anthropic releases Claude Sonnet 5.5, with input and output tokens at half the Opus 5.5 price

Anthropic released Claude Sonnet 5.5 on September 28, 2026, as the second model in the Claude 5.5 family. Its prices match Sonnet 5, and on input and output tokens it costs half as much as Opus 5.5. The model ID on the Claude Platform is claude-sonnet-5-5, and the company lists it on Amazon Web Services, Google Cloud, and Microsoft Azure.

Prices and the savings Anthropic claims

The price list did not move from Sonnet 5. What changed is the comparison with Opus 5.5, which charges $4 per million input tokens and $20 per million output tokens. Cache reads cost $0.20 per million tokens on both models, so a workload dominated by cache reads gains less from the switch than one dominated by fresh input and output.

Anthropic also says Sonnet 5.5 costs up to 30% less per task than Sonnet 5 because it needs fewer tokens for the same work, and that it generates output more than 30% faster. Both figures are the company's own measurements.

• Input: $2 per million tokens, versus $4 on Opus 5.5.
• Output: $10 per million tokens, versus $20 on Opus 5.5.
• Cache reads: $0.20 per million tokens on both models.
• Cache writes: $2.50 per million tokens, versus $5 on Opus 5.5.

Benchmark results, as reported by Anthropic

On Terminal-Bench 4.0, an agentic coding test in the command line, Anthropic reports 70.6% for Sonnet 5.5, 10.3% for Sonnet 5, and 66.4% for Opus 5.5 at Xhigh effort. On CursorBench 4.0 the figures are 55.5%, 34.1%, and 57.8%. On OSWorld 2.1, a computer-use test, they are 80.1%, 57.0%, and 81.8%.

On GDPval-AA v2.1, which scores professional work across 44 occupations, Sonnet 5.5 gets 1844 Elo points against 1846 for Opus 5.5 and 1449 for Sonnet 5. Artificial Analysis ran that test on a pre-release deployment that had a bug affecting structured outputs. Anthropic itself adds that Opus 5.5 remains clearly stronger at complex, open-ended work that needs sustained judgment.

What breaks in existing API code

Thinking is on by default. A request without a thinking field runs with adaptive thinking, and the value disabled, which turns thinking off on Sonnet 5, returns a 400 error on Sonnet 5.5. The replacement is between_tools, the lowest setting, and it is accepted only at low, medium, and high effort.

The migration guide lists four more settings that return a 400 error: thinking budgets, sampling parameters, assistant prefill, and forced tool choice. The context window is 1 million tokens and the maximum output is 128,000 tokens. Default effort is High on the Claude Platform and Medium in Claude Code and the Claude apps.

Safeguards that change which model answers

Sonnet 5.5 is the first Sonnet model launched with cyber safeguards like those on Opus 5.5. Routine bug finding and fixing in a user's own code is unaffected, but higher-risk cybersecurity tasks visibly fall back to Sonnet 5.

It is also the first Sonnet with classifiers meant to stop reasoning extraction. The preserved thinking mechanism ties a model's thinking to the account that created it. According to Anthropic, most developers will not notice, but moving conversations between accounts is affected, including switching accounts mid-session in Claude Code.

Availability

The model is available with zero data retention, as Opus 5.5 and Sonnet 5 are. Anthropic's model overview says Sonnet 5.5 will not be retired before September 28, 2027. The company says Claude Haiku 5.5 will follow in the coming weeks, without giving a date.