Frontier Log

Which Claude Model Switches Keep the Reasoning

Claude Fable 5.2 reads the earlier reasoning of Opus 5.5, Sonnet 5.5 and Haiku 5.5, but only Mythos 5.2 reads Fable 5.2's, Anthropic's API docs say.

A grid of Claude model switches on the Claude API. Rows are the model you switch from: Haiku 5.5, Sonnet 5.5, Opus 5, Opus 5.5, Fable 5 or Mythos 5, Fable 5.1 or Mythos 5.1, Fable 5.2 or Mythos 5.2. Columns are the model you switch to: Fable 5.2 or Mythos 5.2, Fable 5.1 or Mythos 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5. Filled cells say Keeps, grey cells say Drops. The Fable 5.2 column is filled in every row. The Fable 5.2 row is filled only in the Fable 5.2 column. Haiku 5.5 keeps on every switch but Fable 5.1. Sonnet 5.5 keeps on Fable 5.2, Opus 5.5 and itself. Opus 5 and Opus 5.5 keep on Fable 5.2, Fable 5.1 and Opus 5.5. Fable 5 and Fable 5.1 keep on Fable 5.2 and Fable 5.1.
Anthropic's table of model switches on the Claude API, as its preserved thinking docs gave it on Nov 14, 2026. Every model in it keeps its reasoning on a switch to Claude Fable 5.2 or Mythos 5.2. A switch from either of those two to any other model drops it. Download

When a conversation moves from Claude Fable 5.2 to Claude Opus 5.5, the Claude API drops the reasoning from Fable 5.2's earlier turns. A move the other way keeps Opus 5.5's. Anthropic lists each case in its preserved thinking docs. A table there covers ten models and says which switches keep the earlier turns' thinking blocks.

The extended thinking docs state the rule behind the table: "A thinking block is readable only by the model that produced it and certain other models. The API ignores or drops the blocks that the target model can’t read." The docs don't say why each model reads the blocks it does.

Model you move fromTo Claude Fable 5.2 or Claude Mythos 5.2To Claude Fable 5.1 or Claude Mythos 5.1To Claude Opus 5.5To Claude Sonnet 5.5To Claude Haiku 5.5
Claude Haiku 5.5KeepsDropsKeepsKeepsKeeps
Claude Sonnet 5.5KeepsDropsKeepsKeepsDrops
Claude Opus 5KeepsKeepsKeepsDropsDrops
Claude Opus 5.5KeepsKeepsKeepsDropsDrops
Claude Fable 5 or Claude Mythos 5KeepsKeepsDropsDropsDrops
Claude Fable 5.1 or Claude Mythos 5.1KeepsKeepsDropsDropsDrops
Claude Fable 5.2 or Claude Mythos 5.2KeepsDropsDropsDropsDrops

Fable 5.2 reads the other models

Fable 5.2 and Claude Mythos 5.2 read each other’s thinking blocks on the Claude API and Claude Models on Google Cloud (Google managed). There they also read blocks from other models, "including Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5", the thinking docs say. They add: "No other model reads Claude Fable 5.2 or Claude Mythos 5.2 blocks." Anthropic's release notes for Nov 9, the day Fable 5.2 launched, say this includes a move to Opus 5.5 or Claude Fable 5.1.

Fable 5.1 and Mythos 5.1 read each other’s blocks and those of earlier models, except Claude Mythos Preview’s. On the Claude API, they also read Opus 5.5’s. So a switch from Opus 5.5 up to Fable 5.1 keeps the reasoning, and a switch from Fable 5.1 to Opus 5.5 drops it, in the docs’ own example. Opus 5 reaches Fable 5.1 the same way, and so do Fable 5 and Mythos 5.

What Opus, Sonnet and Haiku 5.5 read

Opus 5.5 reads blocks from Opus 5 and from earlier Opus, Sonnet and Haiku models. On the Claude API and Google Cloud (Google managed), it also reads them from Claude Sonnet 5.5 and Claude Haiku 5.5. It doesn’t read them from the Fable and Mythos models, the thinking docs say.

Sonnet 5.5 doesn’t read blocks from Opus 5, Opus 5.5 or any Fable or Mythos model. It reads Haiku 5.5’s on those two platforms. Haiku 5.5 doesn’t read blocks from Opus 5, Opus 5.5, Sonnet 5.5 or any Fable or Mythos model. So a switch from Haiku 5.5 to Sonnet 5.5 keeps the reasoning, and a switch from Sonnet 5.5 to Haiku 5.5 drops it. Neither of them reaches Fable 5.1, but both reach Fable 5.2.

Switching back reads it again

A dropped block doesn’t cause an error, and the API doesn’t bill it, Anthropic’s preserved thinking docs say. The docs’ example moves one history from Fable 5.1 to Opus 5.5 and back: "Back on Claude Fable 5.1, they’re read again, because the API never changes your messages. On a switch, reasoning is lost permanently only if your code removes it."

The switch back keeps the blocks only while the conversation before them stays the same. Fable 5.2 and Mythos 5.2 "enforce the prefix check for every account", Anthropic’s Nov 9 release notes say. Code might edit the system prompt, the tools or an earlier message, then send back a Fable 5.2 thinking block from a later turn. Then "the request returns a 400 error". A dropped block doesn’t count as a change, the preserved thinking docs add, "so both models’ blocks stay valid." A request setting, `prefix_mismatch_behavior`, chooses whether such a request gets the error or the API drops the failing blocks, the preserved thinking docs say.

Models with preserved thinking, Fable 5.2 among them, reject the older `thinking` setting "enabled" with a 400 error. On a move to one of them from a model whose requests set it, the preserved thinking docs say to send "adaptive" instead.

Three conversations drawn as rows of turns. Switch up: an Opus 5.5 turn, an arrow marked Kept, then a Fable 5.2 turn. Switch down: a Fable 5.2 turn, a grey arrow marked Dropped, then an Opus 5.5 turn. Switch back: a Fable 5.2 turn, a grey arrow marked Dropped, an Opus 5.5 turn, an arrow marked Kept, then a Fable 5.2 turn. An arc over the Opus 5.5 turn, marked Read again, runs from the first Fable 5.2 turn to the last.
Our own example of Anthropic's rules. Opus 5.5 can't read Fable 5.2's thinking, so the API drops it for Opus 5.5's turn. Back on Fable 5.2, the model reads its own earlier thinking again, and Opus 5.5's too, while the conversation before them stays the same. Download
Our exampleTurns, in orderWhat the last turn's model reads
Switch up: the reasoning is keptOpus 5.5, then Fable 5.2Opus 5.5’s thinking
Switch down: the reasoning is droppedFable 5.2, then Opus 5.5None of Fable 5.2’s thinking
Switch back: Fable 5.2 reads both earlier turnsFable 5.2, then Opus 5.5, then Fable 5.2Both earlier turns’ thinking

The thinking docs tell developers: "Don’t strip thinking blocks when you switch. The API drops unreadable ones without billing them, and a model that reads them still gets them." Stripping also loses fallback credit, they add. After a refusal, "None of the permitted fallback models can read the refusing model’s thinking", the preserved thinking docs say, and back on the model the developer asked for, Claude reads both models’ thinking.

Platforms and accounts

The table covers the Claude API. The docs name Claude Models on Google Cloud (Google managed) alongside it for the reads involving Fable 5.2, Mythos 5.2, Sonnet 5.5 and Haiku 5.5. The preserved thinking docs add an account rule. Thinking blocks from Fable 5.2, Mythos 5.2, Sonnet 5.5 and Haiku 5.5 work only in the account that produced them or a linked one. In any other account, the API drops them without an error.

Anthropic’s release notes for Sep 1, the day Fable 5.1 launched, said its blocks "are preserved only for the model that produced them or a newer one". Opus 5.5 launched later, on Sep 22, but it doesn’t read them. "A switch from Claude Fable 5.1 to Claude Opus 5.5 drops it", the thinking docs say. The rules for Fable 5.2 came in the notes for Nov 9.

Sources

Models

Related

Tags