The Lab Notes

Anthropic

Claude Sonnet 5.5 keeps Sonnet 5 pricing but breaks five API calls

Anthropic's Claude Sonnet 5.5 charges the same per token as Sonnet 5, but five request patterns now return 400 errors and one change fails silently.

The Lab Notes ·

A brass pocket watch with its lid open, photographed on a plain background.

Claude Sonnet 5.5 costs the same per token as Sonnet 5, but code written for the older model will hit five new API errors, and a sixth change raises no error but silences some apps.

Anthropic released the model in general availability Sept. 28, 2026, according to its model page, which lists it as active and latest. The API ID is claude-sonnet-5-5. Anthropic’s “What’s new in Claude Sonnet 5.5” documentation said it is available on the Claude API to all customers, and on Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry.

Rates match Sonnet 5

Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens, the same as Sonnet 5, including caching and batch rates, the documentation said as of Sept. 28. The Batch API is 50% off. The context window is 1 million tokens and maximum output is 128,000 tokens. Retirement comes no sooner than Sept. 28, 2027.

Anthropic’s Sept. 28 announcement calls Sonnet 5.5 “a clear upgrade over Sonnet 5 that runs 30% faster and costs up to 30% less for most work.” That is Anthropic’s own claim. The per-token rates are identical.

Five requests now return 400 errors

Code running on Sonnet 5 breaks in five places on Sonnet 5.5, Anthropic’s documentation said.

  • Sending thinking of type “disabled” returns a 400. To turn off up-front thinking, developers send “between_tools,” which works at low, medium and high effort and returns a 400 at xhigh or max.
  • Forced tool use is unsupported. A tool_choice of type “any” or “tool” returns a 400, while “auto” and “none” work. Anthropic points developers to strict tool use or structured outputs.
  • Thinking blocks are tied to the model and conversation. Sonnet 5.5 reads blocks from Sonnet 5, Opus 4.8, Haiku 4.5 and earlier models, but not from Opus 5, Opus 5.5 or any Fable or Mythos model, and no other model reads its blocks. On accounts created on or after Aug. 31, 2026, replaying a block after earlier history was edited returns a 400 by default.
  • On the Claude API and Google Cloud, the older computer_20251124 computer-use tool returns a 400. Requests must use the computer_toolset_20260801 toolset. Amazon Bedrock still accepts the older tool.
  • In the advisor tool, which is in beta, a Sonnet 5.5 executor rejects Opus 4.8, Opus 4.7 and Sonnet 5 as advisors with a 400.

One change fails without an error

Apps that stream progress notes between tool calls can go silent on Sonnet 5.5, the documentation said. Text between tool calls longer than a sentence or two now comes back in thinking blocks, and at the default display setting, “omitted,” that text is empty. Turning on between_tools or setting a display value brings it back.

Effort settings need a new sweep

Anthropic’s documentation said the same effort setting no longer produces the same amount of thinking as on Sonnet 5, and told developers to re-run their effort sweep. Default effort on the API is high, per Anthropic’s model page. Anthropic’s Sept. 28 announcement said medium is the default in the Claude apps.

Analysis

We think the unchanged rate card is the least useful number in this release. A team that disables thinking, forces tool calls or streams progress text will see errors or silence before it sees any savings. In our view, Anthropic’s cost claim deserves a test only after the migration checklist is done and the effort sweep has been re-run.