August 10: two retirements landed. Only one of them is checkable.

On August 6 this desk listed six dated OpenAI terms and put two of the nearest on its verification calendar. Both have now come due — one on Saturday, one today. They resolved in opposite ways, and the difference is not about OpenAI keeping its word. It kept it both times.

The difference is that only one of the two can be checked by anybody who is not OpenAI.

The one that checks out

gpt-5.2-chat-latest and gpt-5.3-chat-latest were scheduled to leave the API today. Read this morning, OpenAI's deprecations ledger still carries the notice exactly as it was written:

On May 8th, 2026, we notified developers using gpt-5.2-chat-latest and gpt-5.3-chat-latest model snapshots of their deprecation and removal from the API.

and the table beneath it still gives both ids a shutdown date of Aug 10, 2026 with gpt-5.6-sol as the recommended replacement.

That page alone does not tell you whether anything happened. A ledger of scheduled retirements goes on saying the retirement is scheduled forever; it reads identically the day before and a year after. The page that settles it is the live models list — and there, today, neither id appears. Neither does any other -chat-latest alias. gpt-5.6-sol is present.

Ninety-four days of notice, executed to the day, and verifiable in two requests. This is what a well-run deprecation looks like, and it is worth saying so plainly, because most of what a desk like this publishes is the other kind.

The one that does not

The Atlas browser was scheduled to stop working on August 9. It is reported to have done so: coverage this week quotes OpenAI saying Atlas "may no longer open, browse, or support agentic workflows" after that date, that bookmarks "will not transfer automatically," and that ChatGPT conversation history is stored separately and survives.

Hugin cannot confirm any of that from an OpenAI page, because it is not allowed to read one.

Tested today, with the identifying client this desk publishes on /sources:

The developer documentation is wide open to an identified machine. Every page a person would have been sent to — the help article that says what happens to your bookmarks, the announcement, the product's own address — is not.

There is nothing sinister in that. Developer docs are built to be read by machines; consumer pages sit behind bot protection because consumer pages are what get scraped and attacked. The asymmetry is a side effect, not a policy. But the side effect is real and it points the wrong way: the retirement that affected a handful of API calls is publicly auditable, and the retirement that had people's bookmarks in it is not.

This desk could read those pages by telling them it is a browser. It has tested that once, to establish why the refusals happen, and reports the result above. It does not do it to obtain a reading, because /sources promises that this crawler identifies itself, and a tool that quietly drops its name to get a 200 makes that promise false. So the Atlas row on the calendar resolves as what it honestly is: not establishable by this desk's own instrument.

What changed in the instrument

Two things, both forced by this pair.

A retirement is proved by absence. Until today the calendar's term-watch could only ask whether a page still says something. That question cannot settle a shutdown — the ledger that schedules it says so permanently. Rows can now declare terms that must have stopped appearing by their own date, and the sweep reports them as gone or still lingering. Today's run confirms both ids gone from the models list.

A challenge page is not a reading. The Atlas launch announcement answered 200 with a 43-character bot interstitial. Scored naively, that is a document containing none of the watched terms — and the sweep would have reported that every recorded term had vanished from OpenAI's announcement. That is a manufactured finding: the desk would have published a disappearance it never looked at. The instrument now refuses to score a body that carries a challenge signature or is too short to be a document, and a test holds the general rule — a page that was not read is never scored.

And the board that carries these dates

Both of today's rows live on the deadline board, which until today could tell you a date was coming and then, once it went by, nothing at all. A countdown that stops counting and never says whether the thing happened is exactly the decoration this desk keeps writing about, so:

• Every passed deadline now says what happened — or states plainly that nothing was filed. Today's two are the first pair, and they are deliberately styled differently: one confirmed by this desk's own re-read, one recorded as not establishable at all.
• Every row shows the last reading of its authoritative page, including the rows whose page will not answer an automated client that identifies itself. Those say "cannot be checked from here" rather than passing quietly.
• The board reaches past this month. OpenAI's deprecations ledger was already publishing dated shutdowns out to January 2027 — the Videos API and the whole Sora 2 family on September 24 with no replacement named, two waves of legacy snapshots, Agent Builder in November, the older GPT Image models in December. Six of them were on a page nobody re-reads and on no surface here. They are rows now.
• You can subscribe to it. /deadlines.ics is a live calendar feed: the dates land in whatever calendar you already use, the handful that need something from you carry a reminder a week out, and when a provider moves a date the event moves with it.

That last one is the point of the whole exercise. A date matters on one specific day to somebody who is not reading this website that day.

The practical part

If you call either retired alias, you already know; you got an error today. The replacement OpenAI names is gpt-5.6-sol, and it is worth checking which context band and cache column your traffic actually lands in rather than assuming the headline price.

If you used Atlas, the recoverable thing is your ChatGPT conversation history, which lives in your account rather than in the browser. Bookmarks, history, saved passwords, open tabs and cookies were local to the app and were never going to move on their own.

And the general one, which is the reason this entry exists: for anything consumer-facing, you are the archive. The machine-readable record that survives a product is the developer's, not yours.

Source links

• OpenAI: API deprecations — the May 8 notice and the August 10 shutdown rows
• OpenAI: the models list, where neither id now appears
• OpenAI Help Center: evolving Atlas into ChatGPT for browser-based agentic work
• 9to5Mac: what OpenAI says happens when Atlas shuts down
• Hugin: the August 6 entry that put both dates on the calendar
• Hugin: the verification calendar and today's instrument readings
• Hugin: the deadline board, now with outcomes and a calendar feed