Claude Code: Smarter Model Switching to Smaller Context Windows
Claude CodeClaude Code 2.1.301 changes what happens when a long conversation is moved to a model with a smaller context window: the previous model now summarizes the conversation first so it fits the new one. The release also fixes a case where this switch could spend usage credits on a summary by the previous model without asking the user first. Together, the changes make mid-session model switching more predictable.
Key Takeaways
- Moving to a smaller-context model now triggers a summary by the previous model so a long conversation still fits.
- A bug that could spend usage credits on that summary without asking has been fixed.
- The behavior is a change to an existing workflow, relevant to anyone who switches models mid-session.
- A separate fix stops repeated compaction loops behind LLM gateways that report a different model name than requested.
- The
fableandbestaliases now retry once on the previous Fable model if the newest is unavailable to the organization. - The changes ship in Claude Code 2.1.301, released October 15, 2026.
What changed
Claude Code 2.1.301, released October 15, 2026, improves how sessions behave when a user switches to a model with a smaller context window than the one in use. A conversation that is too long for the new model is now summarized by the previous model first, so the session carries over instead of failing against the smaller limit.
A fix for surprise credit usage
The same release fixes a related defect. In some cases, switching to a model with a smaller context window could spend usage credits on that summary, produced by the previous model, without asking the user first. Users on credit-based billing could therefore be charged for a compaction they never approved. The fix removes that unprompted spend.
Why it matters
Switching models mid-session is a routine move, for example dropping to a faster or cheaper model for a simple follow-up. Before this change, the outcome depended on how large the conversation had grown. The new behavior makes the transition explicit: the earlier, larger-context model condenses the history, and the user is not billed for it unexpectedly.
Related fixes in 2.1.301
The release also stops conversations from being compacted over and over, on an unintended model, when an LLM gateway's responses name a different model than the one requested. It also makes the fable and best aliases retry once on the previous Fable model when the newest one is not available to an organization.