1M-context Opus on the $200 Max plan: what is included and what is capped

The Max 20x plan costs $200 a month. The same page is blunt that this is not an unlimited plan: it grants “20 times more usage per session than the Pro plan,” bounded by “two weekly usage limits: one that applies across all models and another for Sonnet models only,” and Anthropic reserves the right to “limit your usage in other ways.” On paper I am renting a large but finite allowance rather than unrestricted access.

In practice I have not found the edge of it. My normal working pattern keeps several projects open at once, an experiment running, a report being drafted, a proposal taking shape, and I dispatch up to six agents at the same time to keep them all moving. Across that load I have yet to hit either the five-hour session limit or the weekly cap. The allowance is finite; my use of it, so far, is not the binding constraint.

What the subscription includes

The part that changed the arithmetic for me is the context window. Opus 4.8 supports a 1M token context window, and on the Max plan that window is not a metered add-on: the model configuration docs state that “on Max, Team, and Enterprise plans, Opus is automatically upgraded to 1M context with no additional configuration,” billed at “standard model pricing with no premium for tokens beyond 200K,” with usage “covered by your subscription.” I have confirmed this on my own account; the 1M window is included, and a long session bills nothing separately.

The mechanism matters more than the headline number. The help center sizes the standard 200K window at “about 500 pages of text or more,” so the 1M window is roughly five times that in one session. In addition, Claude Code does not let the window fill in practice; it auto-compacts “before the window fills, at about 967K tokens by default” (model configuration docs). The practical effect is that a single agent can hold an entire workstream, the experiment code, the report draft, and the half-written proposal, without my pruning context by hand to keep the session from failing.

Flat cost versus a metered token bill

Put the two together and the reason the plan behaves as if unlimited is specific: the cost is flat, the context per agent is large enough that I rarely restart a session, and my throughput at six agents sits below the caps. Metered against the API (application programming interface) rate of $5 per million input and $25 per million output tokens, a week of that multi-agent load would run well past $200; the subscription converts a variable, and occasionally alarming, token bill into a fixed line item. Consequently, that conversion, not any true absence of limits, is what I am paying for. I looked at the per-token side of the same model in a separate note on Fable 5 versus Opus 4.8, and the flat-fee logic here is the small-scale version of the argument for using the build-out rather than betting on it.

Two caveats

Firstly, “included” is clean on my Max 20x tier but not uniform below it. On the Max 5x tier, in contrast, a reported issue has Opus 1M showing “Billed as extra usage” and, with extra usage disabled, refusing with “not available for your account,” which contradicts the documentation’s “Included with subscription.” It is open with no maintainer response, so I treat 1M-on-subscription as verified for 20x and unsettled for 5x. Secondly, the caps are real even where I have not reached them; a heavier user, or someone on a cheaper tier, will meet them sooner than I do.

At my scale the binding constraint has moved: it is no longer the price or the context ceiling, but whether I can keep six parallel agents coherent in my own head. I have no measurement of that yet. TODO, dated July 17: log a week of actual session hours against the weekly cap, so the next version of this note can replace “I have yet to hit it” with the measured distance to the edge.