Claude Code now assumes a 1M-token context for unrecognised model IDs
Claude Code 2.1.323 raises the context window it assumes for unknown model IDs from 200K to 1M tokens.
Claude Code 2.1.323, dated November 12, 2026, raises the context window it assumes for a model ID it does not recognise, from 200K to 1M tokens, according to the changelog entry. A model ID is the name string a session uses to pick a model. Claude Code assumes a window size for such IDs. Before this release that assumed size was 200K tokens. It is now 1M, five times larger.
The assumed window matters when the real model is smaller. The changelog says that sessions on other vendors' models with a smaller window can fail every turn once full, and keep failing until the user runs /clear. Its advice is to set CLAUDE_CODE_DISABLE_1M_CONTEXT=1 before such a session fills.
The entry gives two ways to deal with the change. Users can run /autocompact to compact the conversation earlier. Or they can set CLAUDE_CODE_DISABLE_1M_CONTEXT=1 to keep the 200K limit. The entry does not say which model IDs count as unrecognised, so a user cannot tell from it whether a given model is affected. The linked docs changelog repeats the same entry and does not add that detail.
Elsewhere in the same release, the Claude apps gateway's /v1/models endpoint now states max_input_tokens for models whose context window is certain on every upstream.
What changed
Claude Code now assumes a 1M-token context window for unrecognised model IDs instead of 200K tokens, which can cause failed turns on other vendors' models with smaller windows.
Type: Breaking change · Takes effect: Thursday, November 12, 2026 · Applies to: Claude Code 2.1.323
Why it matters
Teams routing non-Anthropic models through Claude Code may hit failed turns if their models have smaller windows than Claude Code now assumes.
What you can do
If you run other vendors' models through Claude Code, set CLAUDE_CODE_DISABLE_1M_CONTEXT=1 before those sessions fill up.
How the story developed
Adds the opt-out variable, the /autocompact option and a warning about smaller-window models that fail once full.
Original source
Where we saw it
Claude Code changelog
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works