Under the hoodTier: how much it should matter to you
1Useful: my rating, 1 to 5
2Signal: worth watching, 1 to 5
CompactionArea: what it touches
Internal ChangesKind: in v2.1.281,
Internal ChangesSection of the release
What
The context window is how much conversation the model can take in at once. Auto-compact shortens the conversation automatically when it gets close to that limit. The code that decides the window size for auto-compact now takes a description of the model instead of only its name. The description includes the model's window size, whether it has a native 1M-token window, whether a 1M opt-in is honored, whether usage past 200k tokens is billed, and whether the model is recognized. The labels showing where the size came from (env, settings, clientdata, experiment, model-default, unknown-model, auto) are unchanged.
Why
This is internal restructuring. The finding does not describe any change in when auto-compact happens.
How sure we are
Something disagreesSomething we can check disagrees with this entry, or the writer said they could not settle it.
The writer flagged doubtThe finding does not settle whether the new descriptor fields change the window size chosen for any model.
Anthropic's release notes agreeImproved screen-reader output in /mcp: a disabled server is read as "off" instead of "pending"
