Claude Sonnet 5.5 Ships Five Breaking API Changes for Operators Running Sonnet 5

Five requests that ran without errors on Claude Sonnet 5 will return 400 invalid_request_error on Claude Sonnet 5.5. Anthropic released the new model on September 28, 2026; TheRouter's route for anthropic/claude-sonnet-5-5 was confirmed live the same day, returning a response in 2530 ms on a fresh probe. If your integration has not been audited against the breaking-change list, some of those requests are failing silently right now.

Forced tool use is a hard error

The most common trip point is tool_choice. On Claude Sonnet 5.5, setting tool_choice to {"type": "any"} or {"type": "tool", "name": "..."} returns a 400 immediately:

tool_choice: type "tool" and "any" are not supported for this model.

tool_choice: {"type": "auto"} and {"type": "none"} are still accepted. If your pipeline forced tool selection to guarantee structured output, you need to move to strict: true with strict tool use, or migrate that constraint to structured outputs. For most pipelines, dropping the forced selector and setting the expectation in the system prompt is the lowest-friction path.

thinking: disabled is rejected; use between_tools

Claude Sonnet 5.5 ships with adaptive thinking on by default. To turn it off, send thinking: {"type": "between_tools"} — the model's lowest thinking setting. Sending the old thinking: {"type": "disabled"} returns a 400 that points you to between_tools. If your integration is not using tools at all, between_tools effectively suppresses up-front thinking; the response contains only text, behaving like disabled did on Sonnet 5.

One detail worth noting: between_tools is only valid at low, medium, and high effort. At xhigh or max, a request with between_tools also returns a 400. To run at those effort levels you must use adaptive thinking (omit the thinking field or send {"type": "adaptive"}).

Thinking blocks are model-bound and conversation-bound

Claude Sonnet 5.5 thinking blocks record the model that produced them and are only accepted back in matching contexts. The practical effect: if you cache a conversation that includes thinking blocks and replay it against a different model, or send a block that was produced in a different organization's account, the API either drops the block silently or returns a 400, depending on your account creation date and whether you have sent the thinking-binding-controls-2026-08-01 beta header.

Accounts created on or after August 31, 2026 have strict block-binding enforced by default. A system prompt change, tools update, or earlier-message edit that occurred after a thinking block was produced will cause a 400 on replay. The fix is to keep conversations append-only, using mid-conversation system messages for instruction updates rather than editing the original system field.

For operators switching models mid-conversation — for example, routing a long context to Anthropic on Sonnet 5 and then upgrading to Sonnet 5.5 — the transition is clean: Sonnet 5.5 reads Sonnet 5 thinking blocks. Moving the other direction (onto Claude Opus 5.5 on the Claude API and Google Cloud) also works. Any other cross-model move drops the blocks.

computer_20251124 is not accepted on the Claude API and Google Cloud

Computer use integrations using the older computer_20251124 tool type will receive a 400 on the Claude API and Google Cloud:

'claude-sonnet-5-5' does not support tool types: computer_20251124.

The replacement is computer_toolset_20260801, available on the Claude API and Google Cloud without a beta header. Amazon Bedrock still accepts computer_20251124 on Sonnet 5.5, so Bedrock integrations do not need an immediate update.

The migration replaces the old tools entry with {"type": "computer_toolset_20260801"} and updates the agent loop for member tool_use blocks and toolset_name on results. Integrations already using the toolset, or using the browser use tool, need no change.

Advisor tool pairings are restricted

If you are using the advisor tool beta, a Claude Sonnet 5.5 executor now rejects Claude Opus 4.8, Claude Opus 4.7, and Claude Sonnet 5 as advisors with a 400. Accepted advisors are Claude Mythos 5.1, Claude Fable 5.1, Claude Mythos 5, Claude Fable 5, Claude Opus 5.5, Claude Opus 5, or Claude Sonnet 5.5 itself. Every accepted advisor also returns its advice encrypted as an advisor_redacted_result block — your client can no longer read the advice text directly.

What did not change

Pricing is identical to Claude Sonnet 5: same input and output rates per million tokens, same batch discount, same cache read rate. The models overview on the Claude API places Sonnet 5.5 in the fast-latency tier alongside Haiku 4.5, with a 1M-token context window and 128K max output. Routing operators who need the full specs, including platform availability and retirement date, should consult the Claude Sonnet 5.5 model page directly.

Effort levels have been recalibrated. The same effort value does not produce the same thinking depth as on Sonnet 5, so carrying over a hardcoded effort setting without re-evaluating will alter your results. Anthropic recommends starting at high for general work and medium for agentic coding with well-specified tasks.

Routing to the new model

TheRouter's route for anthropic/claude-sonnet-5-5 was confirmed available on 2026-10-05; a test completion returned in 2530 ms, consuming 26 total tokens (22 prompt + 4 completion). Operators who have their integrations pinned to claude-sonnet-5 rather than the current alias should plan the migration against the five breaking changes above before switching the target model ID. The Anthropic provider page lists the full set of Claude models available through the gateway, and the pricing reference covers token rates for this generation of models.