A context window does not follow the model
An extended window does not appear to carry over automatically from one model to another, a point the documentation does not settle explicitly: the size suffix should be retyped as a precaution at every model change through the /model command, except for two named models whose behaviour differs from the general case.
Switching from one model to another in Claude Code does not automatically carry over an extended context window. The suffix that activates this wider window, such as [1m] for one million tokens, applies to a specific model-selection command and falls back to the standard window as soon as that suffix is not repeated.
The suffix has to be asked for again, it is not inferred
The official documentation describes the syntax of the suffix that activates an extended window, typing /model opus[1m] or /model sonnet[1m] activates the extended window for that specific model, but it never explicitly states what happens to this suffix on a later model change. The rule that follows should therefore be treated as a precaution to verify rather than as a fact documented in black and white: switching to another model through a /model command with no suffix might not preserve the wide window.
/model opus[1m]
/model haikuThe second line carries no suffix. Two cases named by the documentation complicate any general rule on this point: on the Max, Team and Enterprise plans, Opus automatically switches to the one-million-token window with no extra configuration, and on the Anthropic API, Sonnet 5 always runs with a one-million-token window, with no 200,000-token variant nor [1m] suffix to choose. A switch to either of these two models therefore does not behave like the general case described here.
Why this detail weighs on session cost
A session that still believes it has an extended window keeps stacking large files and long tool outputs into it. If the real window has in fact reverted to standard, this practice brings the auto-compact threshold much closer than expected, and the conversation ends up summarised at an unexpected moment. The cost does not come from a different rate depending on the model chosen, it comes from a gap between the assumed capacity and the capacity actually available, a gap that leads to rereading and recompacting more often than it should.
Resuming a session raises the same question differently
Resuming a session with claude --resume keeps the model used at the moment the transcript was recorded, whatever the active model setting is at the time of resuming. This behaviour points the same way: nothing carries over on its own from one choice to the other, the configuration active at that precise moment matters more than the reader's assumed intent.
What still needs to be checked for yourself
User reports sometimes describe the opposite effect, a session that stays stuck on an extended window after a model change that was supposed to revert to the standard window. The right habit is therefore not to guess one way or the other, but to remeasure directly after every change. The neighbouring lesson on the auto-compact threshold shows another place where the model's real window, not an assumed value, remains the only reliable measure.
The window suffix before and after a model change
The extended window suffix is active, the session has the large window announced for the current model.
The new model is selected without the suffix being asked for again; except for a documented exception for that model, the effective window may drop without warning to that new model's standard value.
A trainer activates the extended window suffix on the Opus model at the start of a Claude Code session, on a plan that does not benefit from the automatic extension. An hour later, he switches model to Haiku with the /model haiku command to continue the same conversation.
Write in one sentence what this situation establishes, and in one sentence what it does not establish.
What this establishes: This situation establishes that the trainer switched model mid-session with a /model command that carries no extended window suffix.
What this does not establish: It does not establish what window size now applies to his session, since that depends on a behaviour that must be checked directly on screen rather than inferred from the command typed.
The three most common miscalibrations
- Trop large This situation proves that switching model mid-conversation causes the entire history already exchanged with Opus to be lost.
- Trop étroit This situation allows no observation at all until the exact time of the model switch is known to the minute.
- À côté This situation mainly shows that the trainer's session lasted at least an hour before the model change.
- The suffix that extends the context window applies to a specific /model command; the documentation does not explicitly say whether it carries over to the next model or not, which is a reason to retype it as a precaution.
- Switching model without repeating this suffix might drop the session back to the new model's standard window, except for Opus on a Max, Team or Enterprise plan, and for Sonnet 5 on the API, which do not have this behaviour.
- Resuming a recorded session keeps the model used at the time of that recording, not the default model active at the time of resuming.
- The only reliable way to know the real window after a change is to remeasure it, never to assume it carried over automatically.
After your next model change, immediately type /model to read the window shown, then /context to see how much of that window is already used. If you wanted to keep an extended window, retype the [1m] suffix by hand rather than assuming it followed on its own.
These points depend on an interface or a rule that may have changed since this was written. Check them on your own screen before relying on them.
- The exact direction of any carry-over between models on your own machine: user reports describe the opposite effect from the one documented, an extended window that stays active after a change that was meant to end it. Check with /context right after every change rather than assuming one direction or the other.
Every datable claim in this lesson links here to the public text behind it. A source that does not open proves nothing.
- Anthropic, Claude Code, model configuration, extended context window consultée le 2026-09-02