Anthropic's Opus 5.5 migration guide lists four breaking changes that return 400 errors on requests written for Opus 5

A cheaper model is a free upgrade only if the request that worked yesterday still works today. Amanda Caswell at The New Stack reports that Anthropic's migration guide for Claude Opus 5.5, released on 22 September 2026, lists four breaking changes that return an HTTP 400 error on requests written for Opus 5, plus one quieter change that raises no error at all.

The New Stack notes the price cut: input tokens moved from $5 to $4 per million and output tokens from $25 to $20, while the 1 million-token context window and 128,000-token maximum output stayed the same. That is why the upgrade looks like a model-ID swap on paper.

Thinking, tool calls and model routing all changed shape

The first change removes two thinking controls. The New Stack reports that Opus 5.5 returns 400 when a request sets thinking to disabled or uses enabled with budget_tokens, so effort becomes the only control. Because thinking is always on, responses begin with thinking blocks, so code that assumes the first content block is text has to change. Default effort dropped from high to medium, and Anthropic recommends setting it explicitly and re-running evaluations.

The second change removes forced tool use. The New Stack reports that setting tool_choice to any or tool returns 400, including on the token-counting endpoint, so cost estimates built on those settings will fail alongside the real requests. Anthropic's replacement is auto combined with strict tool use or structured outputs and a prompt that states when the tool applies.

The third change is the one most likely to slip past a router. The New Stack reports that thinking blocks are tied to the model that produced them. Only Fable 5.1 and Mythos 5.1 can read Opus 5.5 thinking blocks on the Claude API, so a fallback to any other model runs the next turn without the earlier reasoning instead of raising an error. Conversations also have to stay append-only, and accounts created on or after 31 August 2026 UTC get a 400 when a client trims old messages, changes tool definitions, summarises earlier context, or rewrites the system prompt mid-conversation.

Computer use and streamed narration

The fourth change affects computer-use agents on the Claude API and Google Cloud. The New Stack reports that Opus 5.5 rejects the computer_20251124 tool and accepts computer use only through computer_toolset_20260801. Each action now arrives as its own tool_use block identified by name rather than by an input.action field, a single turn can carry several of them, and every result has to echo toolset_name. The older tool still works on Amazon Bedrock.

The change most likely to go unnoticed does not throw an error at all. The New Stack reports that on Opus 5, text Claude writes between tool calls comes back as text blocks, and on Opus 5.5 that narration arrives as progress-update thinking blocks. At the default thinking.display setting of omitted, those blocks are empty, so any UI that streams narration to users will go dark between tool calls until developers set display to updates or summarized.

For a team on the Claude API, the honest way to price the upgrade is to run the migration in a development environment first, then measure how many paths hit a 400 or route around invalid thinking blocks. Claude Managed Agents users only need to change the model name, and Claude Code carries its own migrate command. Everyone else pays the price cut back in switching cost.

Source

Amanda Caswell, The New Stack, Anthropic made Opus 5.5 cheaper. Then it broke four things your agent depends on., 23 September 2026.