AI EngineerGuide

Will my thinking blocks survive? Prefix checker for Claude and GPT-6.2

Paste an earlier request and the next one. See the first change before any thinking block, which blocks it invalidates, what each model does (400, dropped or fine) and the documented fix.

Runs in your browser. Nothing you paste is uploaded.

JSON request bodies, or a Python/TypeScript SDK call with literal values. Remove secrets before pasting. Nothing leaves your browser, but your clipboard and screen still can.
Target models

9 reason codes, 4 error strings and 22 fixes, each quoted from an official Anthropic or OpenAI page (last checked 2026-11-20). Reason codes · Error messages

How to use it

  1. Paste the earlier request (or the one that worked) on the left and the next request (or the one that failed) on the right. JSON bodies and SDK calls with literal values both work. For a whole session, tick Or paste a log and paste a JSON array or JSONL.
  2. Check the target models. They’re preselected from B’s model plus every model that runs the check.
  3. Read the verdict per model, the first change (with a mini diff), and the fix. Copy the snippet, or copy the whole report as Markdown.

How the checker works

It works out the API (Anthropic Messages or OpenAI Responses) and compares the two requests the way the provider does. On Claude that means system, tools and every earlier message, as parsed JSON, ignoring cache_control and tools with defer_loading: true that nothing has referred to. It also checks that the thinking blocks you send form an unbroken run. On OpenAI it aligns the input items and checks tools, text.format, reasoning.effort, instructions (before GPT-6.2 only) and safety_identifier.

The first change is the earliest difference before a thinking block or reasoning item. Every block after it, produced before B, is invalid. Then each target model gets its documented outcome:

  • 🔴 400 error: Claude Fable 5.2, Mythos 5.2 and Mythos 5.1 with the setting unset, or any model with prefix_mismatch_behavior: "error".
  • 🟠 Possible 400, or stale thinking billed: Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 with the setting unset.
  • 🟡 Dropped, no error: GPT-6.2 (prefix_mismatch), Claude with "drop_block" (not billed, but it can cost more), and Message Batches items.
  • 🟢 Nothing happens: no change before any thinking, or a model that doesn’t run the check.
  • ⚪ Can’t tell: no thinking in B, history stored server-side, or a change the docs don’t confirm.

What it can’t see: files fetched by URL (the API compares the bytes), MCP tool lists the API fetches, headers your SDK adds, and Gemini (we found no Gemini docs about prefix-bound reasoning). Where the docs don’t say something, the checker says so instead of guessing.

More: what drop_block does and costs, how to build an append-only agent, every reason code and every error message.

How we check this

Every outcome, reason code, error string and fix is data in reasoning_rules.json with a short verbatim quote, its source and the date we checked it. The examples on this page are the same fixtures the tests run. Anthropic pages were last re-read on 2026-11-20; OpenAI pages on 2026-11-19.

Unconfirmed in the docs, so shown as "unconfirmed" or "unknown": whether a string and a text-block array count as the same content on Claude; how the 400 error's position maps when system or same-role messages merge; OpenAI's capability order within a family; reasoning.context defaults on GPT-6.1 Sol and GPT-6.2; whether tools, text.format and reasoning.effort changes remove every reasoning item on GPT-6.2; and switch rules off the Claude API.

FAQ

Why does Claude Fable 5.2 return "The block is bound to a different conversation"?

Something before a thinking block you sent back changed after Claude produced it: system, tools or an earlier message. Common causes are today’s date in the system prompt, MCP tools appearing mid-chat, trimmed history, and serializers that drop empty thinking blocks. Claude Fable 5.2 and Mythos 5.2 enforce the check on every account since November 9, 2026.

Which Claude models run the prefix check?

Claude Fable 5.2, Mythos 5.2, Fable 5.1, Mythos 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5. With prefix_mismatch_behavior unset, Fable 5.2, Mythos 5.2 and Mythos 5.1 return a 400; the others might return a 400, or let the stale thinking through and bill it. Claude Opus 5 and earlier models don’t run it.

Does GPT-6.2 return an error when I edit the conversation?

No. Starting with GPT-6.2, the Responses API silently drops reasoning items after a prefix edit and lists them in input_transformations with the reason prefix_mismatch. Deleting whole items without editing the rest doesn’t trigger it, and top-level instructions are appended after input on GPT-6.2, so changing them isn’t a prefix edit there.

What doesn't count as a change on Claude?

model, effort, max_tokens, tool_choice, metadata, thinking.display and other parameters; JSON formatting and key order; cache_control markers; and a tool with defer_loading: true that nothing has referred to yet.

Is what I paste sent anywhere?

No. Parsing and checks run in your browser, and the page makes no network requests with your input. It isn’t stored or put in the URL. Remove secrets anyway: your clipboard and screen can still leak them.