Claude API documentation: thinking blocks are lost when switching from Claude Fable 5.2 or Claude Mythos 5.2 to another model
Anthropic's documentation describes that thinking blocks from the Claude Fable 5.2 and Claude Mythos 5.2 models are silently discarded when switching to another model within a conversation, without the API returning an error. In the opposite direction — when switching to these two models — reasoning from other models is preserved.
Anthropic has documented how the Claude API handles reasoning blocks (thinking blocks) when a model is switched mid-conversation. According to the documentation, the models Claude Fable 5.2 and Claude Mythos 5.2 can read thinking blocks created by other models in the Claude family — specifically Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5. If the conversation moves to one of these two models, the reasoning from the previous model is preserved.
However, it does not work the other way around: when the conversation moves from Claude Fable 5.2 or Claude Mythos 5.2 to any other model (including Claude Opus 5.5 or the older Claude Fable 5.1), the API discards the thinking blocks created by these two models. According to the documentation, the request nevertheless completes successfully, that is, without an error message warning about the data loss.
According to the documentation, this behavior applies to the Claude API and to Claude Models on Google Cloud, where the service is operated by Google. The documentation refers to a more detailed description in the "Switching models mid-conversation" section.
Why it matters
For developers building applications with dynamic model switching (e.g., a cheaper model for routine queries, a more capable one for complex tasks), it is crucial that the loss of reasoning when moving from Claude Fable 5.2 or Claude Mythos 5.2 to another model does not show up as an error — the API request completes successfully. This means that without knowledge of this behavior, an application can silently lose reasoning context between conversation steps, without this being apparent from the API response.
What was added since the original report
Verified updates
-
Thinking blocks are generated exclusively in Claude Fable 5.2 and Claude Mythos 5.2; No other model reads thinking blocks from these versions; Moving a conversation from Fable/Mythos to another model results in the blocks being dropped; Includes specific models: Claude Opus 5.5 and Claude Fable 5.1
- Thinking blocks are generated exclusively in Claude Fable 5.2 and Claude Mythos 5.2
- No other model reads thinking blocks from these versions
- Moving a conversation from Fable/Mythos to another model results in the blocks being dropped
- Includes specific models: Claude Opus 5.5 and Claude Fable 5.1
Two audiences, two different impacts
What this means
For individuals
Developers who switch between Claude models within a single conversation need to know that reasoning (thinking blocks) is preserved only when moving to Claude Fable 5.2 or Claude Mythos 5.2 — in the opposite direction it is silently discarded, and the API does not report this as an error.
For a business
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.