How this was built,
and where it is weak.

Three kinds of thing are recorded, and they are not the same:

• Reset — usage counters were zeroed for users. This is what the tracker is about.
• Banked reset — a reset was credited to your account for you to spend later. OpenAI shipped this in June 2026. It is not a counter zeroing, so it is counted separately.
• Limit change — the ceiling itself moved, permanently or on a clock. Doubling a 5-hour limit is not a reset. These are hidden by default on the timeline and can be toggled on.

Conflating these is the most common error in coverage of this topic. Selling top-up credits is not a reset. Extending a 2× promotion is not a reset. A reset is the counter going to zero.

Most resets were announced on X, and X returns HTTP 402 to automated fetches — the post bodies cannot be read programmatically. Press reports of these dates proved unreliable: at least one widely repeated date was a full day wrong, and the error had propagated across several outlets.

So dates are not taken from press at all. Every X post ID is a Twitter snowflake, which embeds its own creation timestamp in the high bits. Decoding it gives the exact UTC instant the post was published, independent of anyone's reporting:

timestamp_ms = (post_id >> 22) + 1288834974657

This is how the most recent Anthropic reset is dated 9 July 2026, 18:01:18 UTC and not 10 July as several outlets reported. A Hacker News thread of users noticing their counters reset was created at 18:03:12 UTC — two minutes later — which independently corroborates it.

Each reset is filed under the reason the vendor gave, in the vendor's own words. When no reason was given, it is filed as goodwill and explicitly flagged "No reason given" on the card. A reason is never inferred from press speculation and presented as fact.

Where commentary widely attributes a cause the vendor did not state, that attribution goes in a note, not the classification. The clearest case: the 1 June 2026 Anthropic reset is universally blamed on Opus 4.8's parallel subagents, released four days earlier. Anthropic's announcement says only "an issue that caused some Claude Code sessions to spawn excessive parallel subagents." It names no model. So neither do we.

• Anthropic publishes no usage quantities. Not tokens, not messages, not hours — only relative multipliers. This is why the value model prices quota cycles rather than tokens. Any site quoting you a tokens-per-week figure for Claude Max is guessing.
• ChatGPT consumer prices are Tier B. OpenAI's pricing pages return HTTP 403 to automated fetches. The $20 / $100 / $200 / $25 figures come from press reporting, not from a page that could be read directly.
• The window starts 2 December 2025. That is the earliest reset for which a primary source was found, not necessarily the first one that ever happened. Both companies ran usage limits well before that date. Earlier resets may exist and are not recorded here.
• Tier A includes staff accounts. Most Anthropic resets were announced by @ClaudeDevs and OpenAI's by @thsottiaux, the Codex lead, rather than on a company page. These are treated as vendor-official. Reasonable people could tier them lower.
• Post text is quoted from search-engine indexes. Because X blocks fetching, announcement text was captured from indexed post titles and cross-checked against press and ResetRadar. Where a quote was captured truncated it is marked on the card and left incomplete rather than reconstructed.
• Two undated promotions are missing. Claude's off-peak 2× promotion around 13–28 March 2026, and the Codex app's launch-window 2×, are both real and both excluded, because no dated vendor-hosted source could be retrieved for either.
• "Applied to everyone" is often unverifiable. Vendors announce global resets. Whether they land globally is a separate question that only users can answer. Two events are flagged where they demonstrably did not.

ResetRadar by Giulio Capecchi is the closest thing to this that already existed, and it is good: an hourly poller over the vendors' official accounts, publishing a source-linked ledger, RSS feeds, and a Telegram channel. The research here was seeded from its events.json, and it surfaced roughly fifteen events that a from-scratch search missed.

This tracker diverges in four ways. It classifies the reason as structured data rather than storing the announcement as prose. It prices the resets. It corrects ResetRadar's automated classifier where the automation is wrong — a post announcing the absence of a reset, an export-control notice, and one reset double-counted as two. And it adds at least one event ResetRadar's poller never saw: the 1 April 2026 precautionary Codex reset.

Everything else in this space answers a different question. OpenUsage, CodexBar, the various menu-bar monitors and the AI Limits & Reset Tracker iOS app all tell you when your quota next refills. None of them keep a history of what the vendors handed back, or why.

The dataset is a single JSON file, served at /data/events.json. Every event carries its sources and its tier. If an event is missing, misdated, or misclassified, that file is the thing to argue with — in public, on the issue tracker.