Will Codex reset today? How our forecast works, and why waiting longer doesn't raise the odds

Disclosure: we build Codex Reset. This post explains how its reset forecast works, including where it falls short. The details come from the site's public method page and its live forecast API.

The Codex Reset homepage answers one question first: will Codex reset today? It shows two numbers, the chance of a global reset in the next 24 hours and in the next 48 hours. People read those numbers in different ways. Some treat them as a countdown. Some assume the chance must be climbing because the last reset was a while ago. Neither is how they work.

The last post covered the 5-hour and weekly windows and the kinds of reset OpenAI describes. This one is about the global resets that arrive without a schedule, and how we put a number on them.

Two inputs, kept apart

The forecast has two sources, and it doesn't blend them.

The first is what Tibo (Thibault Sottiaux, @thsottiaux, who leads Codex) posts on X. Almost every global reset so far has been announced there. When he explicitly promises a future global reset, the site opens a ⚠️ Watch. It scores 93% when his words give a date, a range or a clock time, and 83% when they don't. Those are the only two scores. Hedged or suggestive wording can stay visible on the site as context, but it never sends a notification. Posts by other people can't create a Watch, and neither can a follow-up that only adds a time.

The second input is history. When no scored post is live, the site falls back to a cadence model built from the verified record of past resets. That record starts in September 2025 and held 60 global resets when this was written.

The cadence model

The current version is called rate-v4. It takes the last eight completed intervals between resets, plus the time since the most recent one, and estimates how often resets are happening right now. The 24-hour and 48-hour numbers come from that rate.

The part that surprises people is how it treats the current wait. In the method page's words: "The ongoing wait adds exposure without adding a reset, so the baseline decreases as waiting continues." Every quiet day is a day of watching with no reset. The model counts it as evidence that resets are rarer at the moment, not as a sign that one is overdue.

We tested the intuitive version too. The homepage states the result: "models that treated a reset as becoming 'due' scored worse in the backtest." So the number doesn't go up while you wait.

Here is what that looked like on November 5, 2026 at 15:06 UTC. The last verified reset was on October 28, 8.4 days earlier. All of the last eight completed intervals were shorter than that: the typical wait was 2 days, the average 2.1 and the longest 3.5. If waiting made a reset "due", this would be the moment for a high number. The forecast said 32% for the next 24 hours and 54% for the next 48. We read it again at 23:40 UTC the same day, after 8.7 days of waiting: the 24-hour number was still 32%, and the 48-hour number had slipped to 53%.

The forecast card on codex-reset.com (screenshot, November 5, 2026). The historical peak is shown in the visitor's local time; this browser was on GMT+8, so 12 PM–3 PM is 4–7 AM UTC.

Two more rules shape what you see. With no signal live, the page shows at most 46% for 24 hours and 71% for 48 hours; the model's own figures stay in the live JSON. And each percentage is rounded to a whole point and comes with a range: 17–52% for that 32%. The range shows how thin the evidence is. The method page is careful to add that it is "not a promise that the truth lands inside it."

"The cadence model, honestly" on codex-reset.com/forecast-method (screenshot, November 5, 2026). The boxed backtest note at the bottom is the page's August snapshot; the current scores are in the next sections.

What we tried and left out

When Tibo hints at a reset without promising one, it is tempting to nudge the number up. We tested that. The site still shows the state of his reset-related posts over the last seven days, with the post that decided it, but that state isn't counted in the percentage. A test on September 9, 2026 found that weighting the percentage by it made the forecast worse. It stays out until the record shows it predicts something.

At the snapshot above, the state was a tease from an October 31 post. The API marked it "effect": "informational_only" with a multiplier of 1.

How good is it?

The forecast is scored with a walk-forward backtest: each past prediction uses only the data that existed at that moment, and is then compared with what happened. The score is the Brier score, the average squared gap between the predicted probability and the outcome (1 if a reset came, 0 if it didn't). Lower is better.

At the snapshot, over 369 predictions:

• the live model: 0.130
• a naive baseline: 0.133
• the previous model, rate-v2: 0.129

So the live model beats the naive baseline and is slightly behind rate-v2. The differences are a few thousandths. The site says this openly: the API marks the forecast "confidence": "low" with the note "Experimental: walk-forward backtest has not beaten both baselines yet.", and the homepage labels it "Experimental · not above baseline".

The ⚠️ Watch scores have a separate record. The method page keeps a table from before the current alert rules, dated August 27, 2026, exactly as observed: dated commitments had resolved twice, with one hit, and the other bands are marked "insufficient history". Two cases are not a track record yet.

The time of day

Verified resets cluster in some hours. In the record, the most common window is 4 AM to 7 AM UTC. The method page calls this "a rhythm, not a schedule: historical clustering cannot name the next arrival." It tells you when to look, not whether a reset comes today.

What it can't tell you

• Your own limits. Codex Reset never reads anyone's OpenAI account, quota or usage. Your 5-hour and weekly windows are in Usage settings or /status.
• What OpenAI plans. A forecast "is not an OpenAI schedule, leak, or commitment, and it does not predict outages."
• Exactly what Tibo meant. His posts are read by an automated language model, which can get it wrong when he is vague.
• Whether a reset reached you. OpenAI says "'Global' does not mean every account or every usage limit."
• Hard bounds. The range beside a percentage "is not a calibrated coverage guarantee".

Reading the numbers

• With no ⚠️ Watch live, the 24-hour number is a base rate. A 32% day is more likely to end without a reset than with one, so if your weekly usage is gone, plan around the weekly rollover rather than a rescue.
• A ⚠️ Watch at 93% means Tibo gave a time. It is still a forecast.
• ✅ is different. It comes when the Pro and Plus accounts we monitor actually show the reset; Tibo's words alone never send one. If the monitors can't take a reading, a ✅ that Tibo has confirmed says "Our monitor hasn't measured it yet". The channels that deliver these are in the post on alerts.

If you want the inputs behind today's number, the full forecast is public JSON at /api/forecast, no key needed, with the method explained on the forecast method page.

Codex Reset is an independent community tracker and is not affiliated with OpenAI. "Codex" is a trademark of OpenAI.