What a Codex Reset Is Worth

Tibo's reset posts draw two kinds of replies: thank-yous for the free usage, and people asking why their weekly refill just moved to next week. Both are half right, and the difference is one line of arithmetic.

Tibo (Thibault Sottiaux, who works on Codex at OpenAI) is running a 28-day challenge: every day, OpenAI either ships something that makes Codex clearly better, or resets usage limits on every paid account. Thirteen days in, the button has been pressed five times, with a banked reset dropped into every paid account on top.

1. A reset is worth the share of your weekly limit you've used, minus the share of your week that's gone. If you never hit your limit, it's worth nothing.
2. At the challenge's pace, a reset every 2.68 days, a heavy user can get up to 2.6 weeks of usage out of every week.
3. One press costs OpenAI about $11M of compute, most of it on Pro accounts. Every guess is a slider below, so bring your own.

The challenge so far

The five presses have landed 2.68 days apart on average, about twice as often as resets did from August through early October. Day 2 shipped four things and still reset, because 76% of 74,565 voters in Tibo's poll asked for one.

How a reset works

Plus and Pro plans have a weekly usage limit. Pro 100 gets 5x Plus, Pro 200 gets 10x (subscribers from before September 29 keep 20x through October 29), and Pro 500 reportedly gets 25x. At the end of your week, usage drops back to zero and whatever you didn't use is gone. Call that your refill.

A reset also drops your usage to zero, but it starts a fresh 7-day week from your next request, so your next refill moves to 7 days out. That's the catch, and OpenAI support confirmed it for global resets too.

One reset, one subscriber

A reset gives back what you've used. It also pushes your refill back by however much of your week has passed: 5 days into your week, the refill that was 2 days away is now 7 days away. Five extra days of waiting cost 5/7 of a week of usage, so:

value of a reset = share of your limit used - share of your week gone

The answer comes out in weeks of your plan, and on Pro 100 a week costs $23.08:

Maxed out with a day to go, the free week is worth $3.30: all it did was move your refill up one day. On pace to use exactly your limit by the end of the week, it's worth nothing. Saving 70% for a weekend sprint, it costs you $9.56, which explains replies like "So I gain 1% usage and the next reset date pushes 2 days later" (murica1stmuslim). Burn through your whole limit on day one, and it's close to the free week the thank-yous are about: $19.78.

If you never hit your limit, a reset swaps usage you weren't going to spend for other usage you weren't going to spend, so it's worth $0 whenever it lands.

Resets that keep coming

Each reset restarts the clock, so count your week from the last one. If you hit your limit before each reset lands, a reset 2.68 days after the last one is worth 1 - 2.68/7 = 0.62 weeks.

At one reset every 2.68 days, you can burn through at most 2.6 weekly limits a week. If that pace holds, someone who would use three times their limit gets about 10.4 weeks of usage out of a four-week challenge.

One press of the button

A press only costs OpenAI something for accounts that would have run out. For an account that would use D times its limit, what a reset is worth depends on when it lands:

• A surprise reset at a random point in the week is worth (1 - 1/D)/2 weeks. That's under half a week even for the heaviest user, because on average it lands halfway through the week.
• A banked reset used the moment you hit the limit is worth 1 - 1/D, twice as much.
• A reset at the challenge's pace lands 2.68 days, or 38%, into your week, so it's worth at most 0.62 weeks, and only to someone who would use 2.6 times their limit or more.

If the accounts that run out would use 2x their limit on average, a press at the challenge's pace is worth about 0.31 weeks apiece. The rest is multiplication: how many accounts run out, what a week of their plan is worth at API list prices, and what share of list price the compute costs. My guesses:

• 12M paid accounts use Codex or ChatGPT Work in a week (range 6-24M). That's 30% of the 40M active users OpenAI reported across both on October 7.
• 10% are on Pro (5-15%): half on Pro 100, 40% on Pro 200 (all at the old 20x), and 10% on Pro 500.
• 7% of Plus accounts run out (4-15%).
• 30% of Pro accounts run out (15-50%). People buy Pro to use it.
• They would use 2x their limit on average with no limit (1.5-3x).
• A Plus week is worth $66 at API list prices ($55-81), going by measurements from FinOps LLM and InventiveHQ, with Pro tiers at their multiples.
• Compute costs 10% of API list price (5-30%).

Multiplied out, one press costs about $11M of compute, $110M at API list prices, or 92 cents per paid account. Pro is 10% of the accounts and 86% of the bill, and Pro 200 at the old 20x is about half of it on its own, so when that drops to 10x after October 29, a press gets about a quarter cheaper (the receipt has a switch). The five presses so far add up to about $51M, a little under five times $11M because the gaps between them weren't even. A surprise press would cost $7.2M, and the Day 3 banked reset up to $14.5M if everyone who runs out used it right at the limit.

Slide every guess low and a press costs $0.4M; slide them all high and it's $270M. The cost of compute moves it the most, so that's the slider to argue about in the replies. Either way, a press costs somewhere between six figures and nine, and my guesses land on eight.

What the receipt leaves out

• Peak load. If the GPUs are already paid for, the tokens are cheap, and the expensive part is everyone who ran out coming back within the same few hours. Tibo has described a 5-hour limit on Pro as a way "to smoothen the demand curve"; a press does the opposite. When there's no spare capacity, every token given away is a token not sold on the API, and the real cost is closer to the $110M at list prices.
• Strategy. Some people treat the announcement as a deadline: "You've got a couple of hours to use up your limits" (alexgetmancom). Spending down before a press lands makes it cost more than this model says.
• The point of it. Most of what a press gives away goes to the heaviest users, and some heavy users say they're leaving anyway, like the one moving "All 3 Pro x20 plans and the $1000/month API bill" to Anthropic (Entinode). Whether a press keeps people like that decides whether it's worth $11M.

If you're still holding the banked reset, spend it the moment you hit the limit, unless Tibo looks ready to press the button.

As for Tibo: you have the real numbers, and the sliders are right there.