Why an AI’s Reasoning Can’t Follow You to Another Model

Every reasoning model produces two things per turn: the answer you see and the reasoning it used to get there. For a year, the practical assumption in multi-model systems was that both were just messages, and that a conversation could be handed from one model to another with its history intact. As of September 1, 2026 that assumption is false at three of the six vendors in this census, in three different ways.

Anthropic’s Claude Fable 5.1 signs its thinking blocks so that only the model that produced them, or a newer one, can read them; a fallback to a cheaper Claude drops them silently. OpenAI’s Responses API persists reasoning items only within a model family, so GPT-5.6 Sol, Terra and Luna share reasoning and GPT-5.5 gets none of it. Google requires thought signatures to be resent unchanged and says its backend handles compatibility when you switch. xAI encrypts reasoning, DeepSeek returns it as plain text and fails loudly if you lose it, and Meta never returns it at all. Our post on the Claude downgrade trap covered one vendor. This is the table for all six, from each vendor’s own documentation, for anyone who builds a router or a fallback chain.

Anthropic’s thinking documentation states the rule in one line: “A thinking block is readable only by the model that produced it or a newer one, and the API ignores or drops the blocks the target model can’t read.” The full list follows. Claude Fable 5.1 and Mythos 5.1 read blocks produced by each other, by Opus 5, Fable 5 and Mythos 5, and by Opus 4.8 and earlier Opus models, the Sonnet models and Haiku 4.5. “No model other than these two can read a block produced by Claude Fable 5.1 or Claude Mythos 5.1.” A dropped block is removed before the prompt reaches the model, does not count toward input tokens and is not billed; the older model “reasons again from the visible conversation.” If the conversation later returns to Fable 5.1 with the same history, its own blocks are readable again.

The drop is silent unless you send the thinking-binding-controls-2026-08-01 beta header, after which responses carry an input_transformations array naming each dropped block with reason model_binding_mismatch. Fable 5.1 adds a second condition that Mythos 5.1 does not check: the block is also bound to the exact conversation prefix, which is the subject of our framework census on history rewriting. For routing, the consequence is the direction rule: Opus 5 to Fable 5.1 keeps reasoning, Fable 5.1 to Opus 5 loses it, and Anthropic’s refusal fallbacks, which target Opus 4.8 and Opus 5, are downgrades by construction.

OpenAI’s reasoning items are opaque, and the API never returns their text. They persist three ways: through previous_response_id, by attaching a response to a conversation, or by replaying the complete output history yourself; in stateless mode, which applies when store is false or the organisation is under zero data retention, each reasoning item carries an encrypted_content field you pass back. A reasoning.context parameter then controls how much earlier reasoning the model may use: current_turn, the default on earlier models, or all_turns, which the GPT-5.6 family supports and uses by default.

The binding is by family. “Persisted reasoning can be reused only within the same model family. For example, gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna can reuse each other’s reasoning, but reasoning does not carry between the GPT-5.6 and GPT-5.5 families. When you switch model families, the API omits incompatible reasoning from the model’s context, even when reasoning.context is all_turns.” What the API tells you is the effective mode in the response’s reasoning.context field, which OpenAI says to check on every response; the docs we read do not describe a per-item report of what was omitted. For function calling the older rule still applies: pass every item between the last user message and your function output back untouched, even if you trim elsewhere.

Gemini’s thought signatures are “encrypted representations of the model’s internal reasoning” that are “required to maintain reasoning continuity across multi-turn interactions.” In the Interactions API’s stateful mode, with store set to true and previous_interaction_id on later turns, the server manages the thought blocks and “you do not need to do anything regarding signatures.” In stateless mode the rules are capitalised: you MUST always resend all thought blocks exactly as received and should NOT remove or modify them.

On switching models, Google’s guidance is a single sentence: “When switching models within a session, you should still resend the previous model’s thought blocks. The backend manages compatibility.” That tells you what to send and not what you get. Whether a Gemini 3.8 Flash signature carries reasoning into an older Gemini model’s turn, or is quietly ignored, is not stated in the documentation we read, and there is no documented signal either way. Our row records that as “not stated” rather than guessing. For Gemini, the workaround is the stateful mode, which removes the question from your code entirely.

The census reduces to four rules for anyone routing across models mid-conversation. They are stricter than the ones in our cost-routing playbook from earlier in the year, because the vendors changed the ground.

None of this argues against multi-vendor systems. It argues for routing at task boundaries rather than mid-task, for an explicit direction rule per vendor, and for logging the one signal each vendor gives you. That is the discipline our AI transformation practice now builds into every router it ships, and the reason our cross-vendor effort ladder matters more than it did in July: effort is the lever that changes cost without changing the model.