OPEN SOURCES. CLEAR BOUNDARIES.
Codex reset sources & methodology
What we know, how we know it, and where the uncertainty starts.
Public records, with their original context.
We sync the public Codex Resets API every five minutes and keep a local copy. Each record retains its source, timestamp and announcement text. This is a single upstream collection; linking an original post does not mean we independently verified every event.
Historical records use announcement or first-observation times in UTC. Expected reset dates have their own precision and timezone labels. Actual rollout times and account eligibility can differ. Your private account usage is never accessed by this site.
Explore the upstream APIDifferent records mean different things.
Regular reset
A reset announcement or a community observation, as classified by the source. These records power the elapsed timer and gap statistics.
Reset credit
A banked reset credit announcement. It does not show that a reset has already been applied, so it is excluded from the regular-reset timer and statistics.
Announced plan
An explicit upcoming announcement awaiting execution evidence. Even after its scheduled time passes, it remains pending until the source records execution.
Public signal
A time-limited, AI-classified watch from Codex Resets. It can be wrong. Expired signals are hidden, and speculative signals do not trigger push notifications.
A small sample is still a small sample.
We calculate gaps between consecutive regular-reset records. The median and average use all available positive gaps, excluding reset credits and future records.
For the 24-hour comparison, let e be the current wait in days. We count finished historical gaps at least as long as e, then count how many ended by e + 1 day. Their ratio is the empirical frequency.
95% Wilson interval: 0.9%–23.6%. Describes sampling uncertainty only; it is not a validated forecast interval.
We withhold the percentage below eight comparable gaps and label every displayed estimate low confidence. The data may be incomplete, multiple posts can describe a single rollout, observations are not necessarily independent, and reset policy can change. We have not validated forecast accuracy. This frequency does not establish the chance of a future reset.
A notification, without another account.
Browser notifications are opt-in. We store the push endpoint, encryption keys, language and a hashed management token to deliver and remove your subscription. No email, password or OpenAI credentials are collected. Unsubscribing removes the saved subscription and its delivery queue.
Only newly collected reset or credit records from the last 24 hours can trigger alerts. Initial imports and older backfills are silent. The sender retries transient failures up to five times and removes expired subscriptions. Delivery timing depends on your browser, device and network.
Claude records: synced, then annotated.
We sync the public Claude Resets API every five minutes. It is a hand-checked, MIT-licensed collection of announcements from @ClaudeDevs, Anthropic staff and official help pages. On top, we keep our own annotations: why each reset happened per its source, how we verified it, and saved resets that the upstream list no longer shows. Where the two disagree, the upstream facts win and we review the difference.
Only applied resets for everyone feed the timer and gap statistics. Partial resets and saved resets are shown but excluded. Saved resets only take effect when you redeem them, so they are not counted as resets.
We backtested 2026 signals against a 20.7% base rate of any Claude reset in a random 72-hour window. Flagship model launches (Fable 5, Fable 5.1, Opus 5.5) were followed by a reset 3 of 3 times; other launches were not. Codex resets (17%), status-page incidents (22%) and Fridays (14%) showed no useful signal. The sample is small and the reset policy has shifted toward saved resets since September 22, 2026.
Explore the Claude Resets API