The headline scores

Anthropic shipped its second Claude 5.5 model on September 28, six days after Opus 5.5 opened the family. The positioning is blunt: Opus handles complex, open-ended work that needs sustained judgment; Sonnet 5.5 takes well-defined everyday tasks; a Haiku 5.5 for high-volume cheap runs arrives in the coming weeks.

The number Anthropic wants quoted is Terminal-Bench 4.0, a benchmark of real multi-step terminal tasks. Sonnet 5 scored 10.3%. Sonnet 5.5 scored 70.6%. On the same settings, Opus 5.5 scored 66.4%. It is the biggest single-generation leap in the Sonnet line, and the first Sonnet to carry the cyber safeguards and anti-distillation classifiers Anthropic previously reserved for its most capable models. It is also over 30% faster at output generation, available on the Claude Platform, AWS Bedrock, Google Cloud, and Microsoft Foundry, and landed in GitHub Copilot the same day.

The cost claim, split open

The sticker price did not move: $2 per million input tokens and $10 per million output, exactly half of Opus 5.5's $4 and $20. Anthropic adds that Sonnet 5.5 needs fewer tokens to finish the same work, cutting cost per task by up to 30%.

That is where the independent numbers complicate the pitch. Artificial Analysis ran both models on the same harness and found that at max effort, Sonnet 5.5 burned roughly 60% more output tokens per task than Opus 5.5. The savings live at the default and medium effort levels, where Anthropic's own data shows Sonnet 5.5 beating Sonnet 5's best scores at around a tenth of the task cost. Push the model to its ceiling and the discount erodes; on output tokens alone, the reported 60% increase narrows Sonnet's cost advantage to roughly 20%.

The practical read: Sonnet 5.5 is the right default for routine coding, documentation, and office work at scale. For complex open-ended work, both Anthropic and independent testers still rank Opus 5.5 higher. Whether Opus becomes cheaper overall at max effort depends on input and tool-use costs that the headline output-token comparison does not show.

The migration clock

Sonnet 4.5 was deprecated on September 30 and retires November 30, with Sonnet 5.5 as the named replacement. This is not a drop-in swap. Thinking is on by default, which changes both cost and output shape. Five breaking changes to plan for: no disabling thinking, no forced tool use, no replaying thinking blocks across accounts, the old computer tool is gone from the API and Google Cloud, and older models can no longer act as advisors.

Anthropic's deprecation page, as of October 1, lists Sonnet 4.5 but not Sonnet 5, so Sonnet 5 users get a longer runway. Either way, third-party migration notes converge on one instruction: re-test your effort settings and compare real cost on real traffic before November 30. The number that matters is the one your settings produce, not the one on the pricing page.

Sources

  1. [1] PonderoRead source
  2. [2] Codersera complete guideRead source
  3. [3] FoneArenaRead source
  4. [4] CryptoRank / BeInCryptoRead source
  5. [5] DigitalMatters migration FAQRead source