DEV Community

Cover image for OpenAI Redrew One Web Page. Our Parser Read Zero Models.
Nabil A
Nabil A

Posted on AI-assisted

OpenAI Redrew One Web Page. Our Parser Read Zero Models.

On Nov 12, 2026, OpenAI redesigned its deprecations page. Where there used to be a section per notice, there was now one table of upcoming deprecations.

The next time our parser read it, it found 0 models.

Not "fewer models". Zero. A page that had been feeding us retirement dates and replacements was, as far as our code could tell, empty.

1. The part that went right

Since tracks when AI models are retired on each platform, among a lot of other things. Those dates come from the providers' own pages, and for each page we had a hand-written parser that knew its layout.

When a page changes shape, a parser like that does not crash. It just reads nothing. That is the dangerous case, because "I found no deprecated models" and "every model was removed" are one bad assumption apart.

We had a guard for exactly this. A page that suddenly reads far below its usual count is refused, and everything we knew from it stays as last read. So nothing was deleted and no dates vanished.

The guard did its job. It also meant that everything we knew about OpenAI was now frozen.

2. The part that went wrong, slowly

Frozen is fine for a day. It is not fine when the page is still changing underneath you.

Every scan showed OpenAI's source as failing. And the replacements went stale: OpenAI had changed the named replacement for gpt-5-mini to gpt-6.1-terra, and we were still saying gpt-5.6-terra. Nothing about that is loud. It is just wrong, quietly, on every answer that touches the Dec 11 shutdown.

The obvious move was to fix the OpenAI parser for the new table. Then wait for the next redesign and fix it again.

3. Our first plan

Our first plan was better than that, but only a little. When a page's own parser reads too little, hand the page to a general reader that reads it the way a person would, find the model, its date and its replacement, and keep only what the page proves.

We built it and tested it on OpenAI's page. It worked.

That is where I pushed back. OpenAI is not special. Every provider redraws its pages, all the time. A reader that is only tested on the one page that broke this week is tested on the one page we already know about.

So we ran it against every page it could ever be given.

4. Seven things one page could not show us

Run on the other providers' pages, live, the same reader that was right on OpenAI was wrong in seven different ways:

  • It read Anthropic's active models as deprecated.
  • It lost replacements that a page states in a second table, away from the dates.
  • It could not prove Mistral's dates, because two dates drawn side by side came out as one run of text.
  • It missed Bedrock rows longer than the quote it was allowed to keep.
  • It invented bare model IDs on Microsoft Foundry, which writes an ID across two table cells.
  • It took a fine-tuned deployment's date for its base model's date.
  • It dropped the regions that a Bedrock date only holds for.

None of those show up on OpenAI's page. Every one of them would have shipped if we had stopped at the page that broke.

5. The harness came before the reader

What fixed it was not a smarter reader. It was building the comparison first and tuning second.

For every page, the test went like this:

  1. Store what the current, hand-written parser reads today.
  2. Run the new reader on the same live page.
  3. Nothing may move, nothing may appear, and no user may be told anything.

Then the harder half: make some stored rows stale on purpose, a wrong date here, an old replacement there, run it again, and check that exactly those rows get put right. Not more, not fewer.

We also cached the reader's answers by input, so changing the code around it cost nothing to re-run.

6. The rule that made it safe

The other change was about trust. A general reader is flexible, which is the point, and also the risk. So it got a strict job description:

  • It may add to, or move, what a precise parser wrote only on proof that code can check: the fact has to be in a quote, the quote has to be in the page, and the old fact has to be gone from every line about that model.
  • If it says nothing about a row, the row stays exactly as it was. Its silence erases nothing.

That second line is the one I would put on a poster. A reader that cannot find something has not proven it is gone.

7. Then a review found two more

Before release, an independent review went through the change and came back with 19 findings. All 19 were real. 18 were fixed, each with a test. The last one is a stated limit: a page read this way never withdraws a date.

Two of the 19 mattered most. A version's row on Foundry would have been read as the bare alias stored beside it, and production has those rows; our first per-page test had missed it. And a date the old parser could not read, like "December 2026", would have withdrawn a retirement that was still standing.

After the fixes, on Anthropic, Gemini, Vertex AI, Microsoft Foundry, Bedrock and Mistral: 0 rows moved, 0 rows added, nobody told, and every row we broke on purpose came back right. On OpenAI's redrawn page, end to end in our test environment: the source was active again, 166 models read, the 16 replacements that had changed were moved, and nothing was withdrawn.

It went to production on Nov 13 at 5:40 PM MST.

8. The same week

The same week, two providers moved dates. Microsoft Foundry pushed claude-haiku-4-5 and claude-opus-4-5 out to Feb 22, 2027. Google pushed its Gemini text-to-speech previews from Nov 17 to Dec 18, 2026.

That is the whole point, really. These pages are not documents you read once. They are live data that someone redesigns whenever they feel like it.

When this is the wrong call

If you read one page you control, a hand-written parser and a loud failure are simpler and better. A general reader earns its place when the pages belong to someone else, there are many of them, and they change without telling you. Even then it only works with the boring parts around it: the harness, the proof rule, and the guard that refuses a page that suddenly reads as empty.

The thing I took from it is smaller than all that. When a fix works on the case that broke, that is the moment to go find every other case it will meet, before the fix meets them in production.


What's the worst thing a page you scrape changed without warning, and how long did it take you to notice?

I'm Nabil. I build Since, which tracks what your code depends on and tells you before it breaks.

Top comments (0)