If you pay for Codex, the size of your month now depends partly on announcements you cannot schedule.

OpenAI's tracked Codex usage resets now number 70, up from 52 on September 8.

	Sep 8	Today
Tracked resets	52	70
Mean interval	6.98 d	5.94 d
Median interval	3.14 d	2.82 d
Fault given / no fault given	22 / 30	24 / 46

The eighteen new rows sit a mean of 2.99 days apart. Two give a fault, and both needed the thread to see it: "Reset all propagated. Sweet dreams." followed an update on quality issues, and "Reset all propagated. Enjoy." followed "Apologies for the slow start with GPT-6.1 Sol". No later post we read gives a fault as its occasion.

The pledge

On October 4, OpenAI's Tibo Sottiaux wrote: "Over the next 28 days, each day we'll either ship one thing that is a clear improvement and relevant for most codex/work users or ship a full reset." The independent tracker has kept a daily log, with Day 1 on October 5.

Through Day 27, it shows thirteen days with a reset and fourteen with only improvements. Each of those thirteen matches one row in this desk's list. Day 28 is today. Sottiaux said on Day 26 that the "next two days will be resets… banked", and when the tracker was read this afternoon it showed Day 28 as announced with distribution unconfirmed.

The improvements were real. Day 1 made GPT-6 Astra and GPT-6.1 Sol about 50% faster by default, Day 16 added persistent caching, and on Day 25 Sottiaux wrote: "We are doubling all usage limits for all Pro 100/200/500 tiers and Business Premium, effective today." The same post adds, "We heard your kind feedback and won't be reducing usage limits for Pro 200, so your Work and Codex usage will stay the same." The post does not explain how those two sentences fit together for Pro 200.

The tracker lets visitors vote on each day, and its numbers are a poll of its own audience. Resets drew between 74% and 97% applause. Improvements ran from 5% to 93%. Sottiaux described the dynamic himself on Day 2: "I did calibrate it and it seems that the game is rigged in reset's favor, but such are the rules at the moment."

A free reset is worth having, and I am not arguing that OpenAI should give out fewer of them. The problem is what they do to the thing you bought.

What the company's own pages now say

OpenAI's help page on resets sorts them into banked, automatic and purchased. Free banked and automatic resets, it says, are "discretionary promotional benefits, with no guaranteed availability or schedule." They "do not permanently increase your plan's limits."

The purchased kind is a "paid instant reset", offered to "selected Plus and Pro personal accounts". OpenAI's page on buying one is plain about what it is: it "pulls your normal weekly allowance forward rather than adding a separate usage entitlement." A purchase restores usage immediately and starts a new weekly period, and OpenAI says it generally does not refund one.

Put those together. Some subscribers get free capacity at unpredictable moments, and some are offered the chance to pay to spend their own allowance early. The people best placed to benefit from free resets are those watching one person's X account or a tracker, with work flexible enough to move around them. Someone with a fixed deadline, or who only reads the plan page, gets less out of the same price.

December

OpenAI's help center says that "starting in December", Plus, Pro and Business allowances "will be shared across ChatGPT and Codex" and measured in tokens. Use will "depend on the model you choose and the task you ask it to do", though "Instant chat will not draw from your usage." The page says OpenAI expects most users won't need to wait for a reset more often than they do now.

The page read today gives no token figure for any plan and no exact date. A subscriber cannot check that expectation without a before-and-after number. The Day 25 doubling makes the number more important, because Pro and Business Premium subscribers have no published way to tell what that doubling becomes once ChatGPT's non-instant use draws on the same allowance.

What should change

Publish the token allowance per plan before the switch, and date it. A subscriber deciding whether to renew in November should be able to read what they will get in December. Say whether the five-hour and weekly windows survive.

Keep a first-party reset log. A dated list on OpenAI's own site should show which plans and allowances each reset covered, and whether it was automatic or banked. "Global" on OpenAI's page "does not mean every account or every usage limit", and an outsider cannot see the difference. A volunteer tracker should not be the record of what customers received.

State the eligibility rules for paid resets. OpenAI says availability "varies by account, including your plan and billing country." That names two factors and leaves the list open. Customers left out of a paid offer should be able to learn why.

A correction to our September piece

On September 8 this desk wrote that Anthropic's last dated change to Claude usage limits we could point to was a May 6 post, that it had "no announcement account", and that grants since then had been "granted silently". That was wrong on our own record: our August 3 article had already quoted @ClaudeDevs on later dated changes, including a June 20 reset across all plans. It is also out of date. Weekly limits in Claude Code settled 25% above their pre-promotion level from September 14, and Anthropic's September 22 Opus 5.5 announcement raised five-hour limits on several plans and gave subscribers a reset they can save and use when they choose.

A narrower point still stands. Anthropic publishes no complete reset log, so no public count of its grants can be shown to be complete. The operator's two unrecorded observations from September remain unverified.

A Reddit post today, attributing its claims to an unnamed insider, says a new Opus will arrive with a bankable reset. The September grant is a precedent for that and not evidence of a release. Our record on today's model rumors sets out what can be established.

How to check this

Every id is in reset-cadence-data.ts, and each date can be decoded from the id without trusting anyone. The date is the post's, so several rows, such as "Lands tomorrow before 10am PST", sit hours ahead of the reset they announce, and the tracker's displayed time can differ from ours. One row is a reply to another user reading only "Shhhhhh", which the tracker files as a banked reset and we kept on its authority, marked ambiguous. Tell us which rows you would file differently.

Thirteen resets in twenty-seven days, and a paid one that only pulls your allowance forward. Publish the allowance before December.