Claude Sonnet 5.5: the cheaper model doubles Opus 5.5 on the agentic coding test

On 28 September 2026, Anthropic released Claude Sonnet 5.5, the second model in the Claude 5.5 family. The announcement is more interesting than a routine launch: it documents a model that is cheaper and faster, that outscores the most powerful model in the family on the agentic coding test, and that loses on every other published test.

The figures Anthropic published

The honest reading of that table has two levels, and the second matters as much as the first.

On Terminal-Bench 4.0, Sonnet 5.5 beats Opus 5.5: 70.6% against 66.4%, a lead of more than four points, against a previous family generation that scored 10.3%. On every other published test, the order reverses: Opus 5.5 stays ahead on CursorBench, on GDPval-AA and on Humanity's Last Exam. The gaps there are small — two points on GDPval-AA, 1,811 against 1,822 — but they are consistent.

Anthropic does not hide that balance sheet. Its own wording is explicit: Opus 5.5 remains clearly stronger at complex, open-ended work requiring sustained judgment, while acknowledging that on several evaluations Sonnet 5.5 at Max effort performs comparably to Opus 5.5.

Why this is not a replacement

Anthropic positions Sonnet 5.5 as a complement, not a substitute. The release describes it as a faster, cheaper model aimed at well-scoped work: fixing bugs, producing polished documents, slides and spreadsheets. Opus 5.5 remains the model for open-ended work that demands judgment.

That distinction matters to anyone buying. For a coding agent that runs for hours, $4 less per million input tokens and $10 less per million output tokens changes the cost of a long task. For a decision that turns on judgment, Anthropic's own text says the cheaper model is not the choice.

Speed and cost: what the release states

• 30% faster than Sonnet 5 in generation;
• up to 30% cheaper per task than Sonnet 5, because it uses fewer tokens for the same work;
• on several tests, Sonnet 5.5 at Low or Medium effort beats Sonnet 5's best score for about a tenth of the cost.

Availability

Anthropic states immediate availability on all of its platforms, including Amazon Web Services, Google Cloud and Microsoft Azure, with a zero data-retention option. On the roadmap, Claude Haiku 5.5, built for high-volume, cost-sensitive use, is due to join the family in the coming weeks.

The release also mentions a new between_tools setting, which is required when thinking is turned off.

What the announcement does not say

Precision matters here, because the temptation of a single headline number is strong.

• The scores come from Anthropic, not from an independent evaluation. Terminal-Bench is unusually hard to fake — the tests are the arbiter — but the protocol and the results are not audited here.
• No real running cost is given for a long-running production agent, where tokens consumed depend on the task rather than the brochure.
• The context window is not published in this release, so we do not state one.
• The "Sonnet 5.5 System Card" is cited by Anthropic for evaluation detail; it is not reproduced here for lack of a verifiable URL.

Why it still matters

Because the hierarchy "the most expensive is the best" still holds — but it is no longer universal. A model at $2 against $4 per input token outscores the $4 model on one measured test, and loses by two points on another. Both facts are true at once.

The useful reading is therefore not "cheaper wins", but: the capability-to-price ratio is moving to the middle tier, and the buyer has to choose according to the kind of work. To place this against the competition, see our AI model showdown. For what the term "agent" really covers, read our explanation of AI agents.