The arithmetic

You pay for a cap.
Resets quietly raise it.

The sticker price of an AI subscription buys you a fixed amount of usage per week. But if the vendor zeroes that counter ten times in seven months, the number of weeks of usage you actually got is not the number of weeks you paid for. Here is what the difference is worth — and the reasons it might be worth nothing at all.

Why this page counts quota cycles and not tokens. The natural way to build this would be "$200 buys you N tokens; resets make it N × 1.4." That calculation cannot be made honestly. Anthropic publishes no token, message, or hour figure for any subscription tier — only relative multipliers like "5×" and "20× more usage than Pro." Any tokens-per-dollar number for Claude would be invented precision dressed up as fact.

So this model prices the one unit both vendors actually define: the weekly quota cycle. Both meter a rolling weekly window. A reset zeroes it. Everything below is built from an official price, a cited event count from the timeline, and one assumption you control yourself.

Claude prices are from Anthropic's public pricing page. ChatGPT prices could not be retrieved directly (the pricing page blocks automated fetches) and come from press reporting — see gaps.

The dataset begins with the first logged reset and ends at the last verification.

50%

The share of a full quota cycle you actually recovered, averaged across resets. This is the honest knob. A reset returns the remainder of the cycle you were in: capped on day one, you recover nearly everything; capped on day six, almost nothing. And if you never hit your limit, a reset is worth exactly zero. Drag it to 0% to see that.

Effective discount on your sticker price
—

What was counted

    Every plan, same assumptions

    Recomputed live from the controls above. Each plan is compared against its own sticker price.

    Do not read this table as a cross-vendor comparison of value. A Claude Max 20x weekly cycle and a ChatGPT Pro weekly cycle are not the same quantity of compute, and neither company publishes enough to say by how much they differ. What the table shows is how far each plan's own effective price has drifted from its own sticker price. Comparing the discount columns is fair. Comparing the dollar columns across vendors is not.
    PlanStickerResets $ / cycleEffective $ / cycleDiscount

    What this number is not

    It is an upper bound, not a measurement. Every reset is counted as though it landed on an account that could use it. Nobody is capped every week.

    It says nothing about whether the plan was good value to begin with. A 27% effective discount on an overpriced plan is still an overpriced plan. This measures drift from the sticker price, not the merit of the sticker price.

    Roughly two in five of these discounts are compensation for a defect. On the timeline, the largest category of reset by far is incident — the vendor broke something and gave usage back. Counting that as a discount is a little like counting an airline's delay voucher as a fare cut. It is real money, and it is also evidence you were sold something that did not work.

    Announced is not delivered. OpenAI Support has confirmed that at least one publicly announced global reset "did not universally apply across all quota types or account states," and users reported the free banked reset never arriving. The strict-count checkbox exists for this reason.

    Resets are not promised. Nothing in either company's terms commits them to any of this. The cadence could stop tomorrow, and the sticker price would be exactly what you pay.