Home / Context and cost

A context window does not follow the model

The context window is a property of the model: recent models have the large window as their base, two older models ask for it through the [1m] suffix, and every model change is checked on screen with /model then /context.

The context window belongs to the model, not to the session: the one you had on a model does not follow you to the next. On recent models, the large window, the one the [1m] suffix designates, is the base window; on two older models it is obtained through that suffix; and changing model changes the window that applies.

Two families of models

Claude Code's model configuration documentation distinguishes two cases. Fable 5.1, Fable 5, Sonnet 5 and later, Haiku 5.5, and Opus 4.7 and later run with the large window and no suffix at all. Opus 4.6 and Sonnet 4.6 reach it only through their [1m] variant, whose access depends on the plan: included in the subscription for Opus 4.6 on Max, Team and Enterprise, subject to usage credits for Sonnet 4.6, Max included.

/model claude-opus-4-6[1m]
/model haiku

The first line asks for the extended variant of one specific model. The second changes model and carries no suffix: the window that applies is then the chosen model's, to be read on screen. On an alias that already designates a model with a large base window, the suffix changes nothing.

What happens when you change model

Claude Code caps the auto-compact window at the model's own window. Since October, when the conversation is too long for a new model with a smaller window, Claude Code's changelog says it is first summarised by the previous model: a model change can therefore trigger a summary. The /autocompact setting is saved per model, so each model keeps its own when you switch. Behind a gateway (a custom ANTHROPIC_BASE_URL address), models that have the large window use it too; if the gateway stops at 200K, /autocompact 200k brings the threshold within its reach.

Why this detail weighs on session cost

A session that believes it has a large window keeps stacking large files and long tool outputs into it. If the real window is smaller, the auto-compact threshold arrives much sooner than expected and the conversation is summarised at an unexpected moment. The cost does not come from a different rate depending on the model, it comes from the gap between the assumed capacity and the available capacity, which leads to rereading and recompacting more often.

A resumed session keeps its model

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, and therefore that model's window.

What still needs to be checked for yourself

The documentation page is the authority on models and plans, and it changes with every new model: reread it before relying on the table. After every model change, /model then /context give the session's real window. 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.