Changing only the model string to Opus 5.5 or Sonnet 5.5 returns HTTP 400 on five settings

The error text points to the fix. In the author's captured response, the API says `thinking.type.disabled` "is not supported for this model" and tells callers to use "thinking.type.adaptive" and "output_config.effort" to control thinking ^[5]. For Opus 5.5 the author's advice is to omit thinking, which is the same as adaptive, and to lower output_config.effort wherever the code used to disable thinking ^[6]. Probing is cheap. The author ran the requests with max_tokens at 64 and says a 400 is not charged ^[3].

Sonnet 5.5 needs two edits in one call. In the author's side-by-side, the Sonnet 5 request sends disabled with effort xhigh, and the Sonnet 5.5 request sends between_tools with effort high ^[11]. The reason is that between_tools returns 400 at xhigh and max, accepts no other field, and cannot be combined with per-message effort changes ^[9]. A port that swaps only the type string keeps xhigh and still fails ^[23].

"Off" on Sonnet 5.5 is not silent. The short updates the model writes between tool calls come back as thinking blocks ^[8]. Code that reads any thinking block as proof of reasoning will misjudge those. In a request with no tools, between_tools answers without thinking at all, so multi-step reasoning tasks go back to adaptive ^[10].

Forced tool calls need a different pattern. The replacement for tool_choice of type any or tool is auto, strict: true on the tool, and a prompt telling the model when to use the tool ^[13]. A request can carry at most 20 strict tools, and every object in them must set additionalProperties: false ^[14]. On Amazon Bedrock neither model has structured outputs, so callers send auto, name the tool in the prompt and validate the input themselves ^[15].

Replay is the quiet break. Each thinking block is tied to one model and one conversation prefix ^[17]. If you change the system prompt, the tools or earlier turns and then send the blocks back, the result is a 400 on accounts created on or after August 31, 2026, where the check is enforced by default, and a dropped block otherwise ^[18]. In our view that means an agent which rewrites its system prompt mid-history fails loudly on new accounts and loses reasoning context on old ones. The author's remedy is append-only history, with instructions changed through a mid-conversation system message ^[19].

Two more rejections are lookups. Computer use is declared as the computer_toolset_20260801 toolset, and Sonnet 5.5 rejects computer_20250124 on every platform ^[20]. A Sonnet 5.5 executor rejects Opus 4.8, 4.7, 4.6, Sonnet 5 and Haiku 5.5 as advisors ^[21].

This is one author's account. Opus 5.5 shipped on September 22, 2026 and Sonnet 5.5 on September 28 ^[1]. The author says all five settings appear in Anthropic's migration guides but reproduced only the first against the live API, on October 6 ^[24][2]. We would run the list as a test plan on your own account, since the replay check depends on when the account was created. The author also published an open-source scanner, claude5-migration-kit, that reports file and line for each hit ^[22].