Reading
Sonnet 5.5's Scariest Change Doesn't Throw an Error
Four Sonnet 5.5 changes return a 400 error. Two others fail quietly, with no error at all. Here's what to check in your agent loop before you switch.
Prices last read 2026-10-03
Short answer
Yes. Code that works on Sonnet 5 can break on Sonnet 5.5, and two of the changes won't show you any error.Anthropic released Claude Sonnet 5.5 (claude-sonnet-5-5) on September 28, 2026, and lists five breaking changes for code already running on Sonnet 5. Four of them return a 400 error: thinking set to "disabled", forced tool_choice, the old computer_20251124 tool on the Claude API and Google Cloud, and advisor pairings that are no longer allowed. The fifth, thinking blocks tied to the model and the conversation, returns a 400 when you replay a block after editing earlier history. That applies by default to accounts created on or after August 31, 2026. The quiet ones are different. A Sonnet 5.5 thinking block sent from another account is dropped and the request still succeeds. And text between tool calls now arrives in thinking blocks, so a UI that streams it can go silent.Source: What's new in Claude Sonnet 5.5 · 2026-10-03; Claude Sonnet 5.5 model overview · 2026-10-03
The four that fail loudly (good)
- thinking: "disabled" is rejectedIt returns a 400 invalid_request_error that points you to between_tools. Send thinking: {"type": "between_tools"} instead. It only works at low, medium or high effort. At xhigh or max it returns a 400, so use adaptive thinking there.
- Forced tool use is gonetool_choice set to any or to a named tool now returns a 400. The same check applies to the token-counting endpoint. Anthropic's fix is auto plus strict tool use, or structured outputs, and saying in the prompt when the tool applies.
- computer_20251124 is refused on the Claude API and Google CloudOn those platforms, computer use now needs computer_toolset_20260801. Bedrock still accepts the old tool. Moving over means changing your agent loop, not just the type string: member tool_use blocks, batch actions, and toolset_name on results.
- Some advisor pairings are rejectedOpus 4.8, Opus 4.7 and Sonnet 5 can no longer act as advisors for a Sonnet 5.5 executor. Every advisor Sonnet 5.5 does accept returns its advice encrypted, so your client can't read it.
Source: What's new in Claude Sonnet 5.5 · 2026-10-03
Note that some coverage frames this differently. Mixed News, for example, counts account binding among the five breaking changes and lists the advisor restriction separately. Anthropic's own list puts the advisor rule fifth and describes account binding as a separate behavior. Either way, the code you need to check is the same.
Source: Mixed News: Claude Sonnet 5.5 breaking changes · 2026-10-03; Claude API release notes · 2026-10-03
The two that fail quietly (bad)
Sonnet 5.5 thinking blocks only work in the account that created them or in an account linked to it. If another account sends one, the API drops it before the model sees it, the request succeeds, and the model answers without that reasoning. Without the thinking-binding-controls-2026-08-01 beta header, you get no signal. On the Claude API and Google Cloud, adding that header makes the response list each dropped block in input_transformations with reason "organization_binding_mismatch". The same quiet drop happens when you switch models mid-conversation and the new model can't read the old blocks. That includes moving from Sonnet 5.5 down to Sonnet 5 through server-side fallback. In that case the reason is "model_binding_mismatch".
Source: Preserved thinking · 2026-10-03
When another account sends one of these blocks, the API drops the block before the model sees it, and the request succeeds.
Anthropic, What's new in Claude Sonnet 5.5 (read 2026-10-03)Source: What's new in Claude Sonnet 5.5 · 2026-10-03
The second quiet change is the one Anthropic itself flags as altering the response without failing any request. Between tool calls, notes longer than a sentence or two now come back as progress-update thinking blocks. At the default display setting, their text is empty, so an app that streams those notes to users goes quiet. To fix it, set a display value that returns the text, or use between_tools.
Source: What's new in Claude Sonnet 5.5 · 2026-10-03
Our take: a 400 is a gift
Everyone will complain about the forced tool_choice removal, and those complaints are overrated. Your tests catch it in five minutes. The underrated risk is the quiet drop. When an agent loses its earlier reasoning, nothing crashes. It just gets a bit dumber and nobody notices, and you end up blaming the model instead of your plumbing. Anthropic does document the behavior and offers a way to see it, but making that visibility opt-in through a beta header is the wrong default. If your pipeline moves history between orgs, workspaces or fallback models, don't change the model name until you've turned that reporting on.
Before you switch the model name
- Search your code for "disabled" thinking and replace it with between_tools (high effort or lower) or adaptive thinking.
- Remove tool_choice any/tool. Use auto with strict: true, or structured outputs.
- On the Claude API or Google Cloud, migrate computer_20251124 to computer_toolset_20260801.
- Check advisor pairings against Anthropic's list of accepted advisors.
- Keep history append-only. Change instructions through mid-conversation system messages instead of edits.
- Send the thinking-binding-controls-2026-08-01 beta header in staging and log input_transformations.
- Check that your UI still shows progress text between tool calls.
Source: What's new in Claude Sonnet 5.5 · 2026-10-03; Preserved thinking · 2026-10-03
Who this affects: developers using the Messages API on Sonnet 5 or older Sonnet and Haiku models. If you use Claude Managed Agents, Anthropic says you only need to update the model name. In Claude Code, the bundled Claude API skill (/claude-api migrate) can apply the swap and the breaking parameter changes, and it gives you a checklist of things to verify yourself.
Source: Migrating to Claude Sonnet 5.5 · 2026-10-03
Optional: compare Claude subscription offers from sellers (a shopping listing, not official access or an endorsement). That link carries our partner code Sk52g, which is how a purchase is attributed to this site. Promotion code UPD3S.
Sources (all read 2026-10-03)