AI Coding Weekly

Theo says a $200 Claude sub bought $9,000 of Opus usage

His own accounting of his own accounts, not a number from Anthropic.

Theo says he has run three Claude accounts down to 0% since Opus 5.5 shipped, and when he looked at what that usage would have cost he got $2,086, $2,442 and $2,182. He puts the average at $2,200 a week, which he extrapolates to roughly $9,000 a month of Opus usage against a $200 Claude Code subscription.

Theo - t3.gg
@theo
X
I have managed to run 3 Claude accounts down to 0% since Opus 5.5 dropped.
Sep 28, 2026 · View on X

That is a 45x ratio if you take the numbers at face value, and the caveats matter. This is one person's accounting of his own three accounts, priced by him, not a figure Anthropic published. It is also a heavy user's month, someone who burned three separate subscriptions to zero rather than one, so it says more about the ceiling than about what a typical engineer pulls out of the plan.

Why the ratio moved

The number is only interesting because it changed. We reported earlier that Theo did the math on the usage limits and put the jump from Fable to Opus 5.5 at a 4x to 6x increase. Drain three accounts at the new ceiling and the dollar figure at API rates gets large fast.

For working engineers, the practical read is that the constraint on frontier model work right now is the subscription limit, not the per token price. If you are budgeting agent work by multiplying tokens by list price, you are estimating a cost nobody on a subscription plan actually pays. The corollary is that the limit is the lever the labs keep pulling, and it can move in either direction.

It also frames the tasks people have been throwing at Opus 5.5. Theo's four month Rust port that finished in 10 hours is the kind of job that eats a usage budget in one sitting. At API prices that workload has a real number attached. At $200 a month it mostly has a percentage counter attached, and the counter is what tells you when to stop.

Nobody outside Anthropic knows how long the current limits hold.

Get the next one by email

Coding with AI, read daily so you do not have to. The experiments, the receipts and the arguments from engineers shipping real software. Not a changelog.