GitHub's Copilot Docs Contradicted Its Own Changelog on Retirement Day

/research/documentation-drift-model-retirements/

We maintain a model tracker with a seven-day review interval. On 1 September 2026 that interval lapsed, and the re-verification produced a result worth publishing.

What we found

Six models were deprecated across GitHub Copilot on 1 September 2026. The changelog entry announcing it is dated 31 July 2026 and names them:

On the same day, GitHub’s supported-models reference page listed all six as generally available, and contained no deprecation notice, no retirement date and no successor for any model. Its only statement on the subject was the generic line that “model availability is subject to change”.

The model-comparison page went further. Of the six models it placed in the deep reasoning group, two were retiring that day — Claude Sonnet 4.6 and Gemini 3.1 Pro. Of the three in visual work, the same two. The page was recommending models for a task on the day they were withdrawn.

The successor that retires first

One detail is worth separating out, because following the guidance literally produces a wrong outcome.

GitHub names MAI-Code-1-Flash as the replacement for Raptor mini, which was deprecated on 1 September. A second changelog entry, dated 11 August, deprecates MAI-Code-1-Flash on 10 September.

Migrating as instructed means moving twice in nine days. MAI-Code-1.1-Flash is the destination that survives both dates, and neither entry says so — the inference requires reading both.

Why this matters beyond the specific models

The practical consequence is which source you check.

Reference documentation is the natural place to look and is the wrong place to look for this class of fact. It is written to describe a current state and updated when someone updates it; a changelog is written to record an event on a date and is correct by construction. When the two disagree, the dated one is the one to trust.

For anyone maintaining a prompt file, a CI workflow, a runbook or a comparison table that names a model, that means subscribing to the changelog rather than periodically re-reading the reference page.

What we changed as a result

Our model tracker’s source list now includes the changelog entries alongside the reference pages, with a comment recording why. The change history on that page carries all seven retirements with their sources.

We also corrected something of our own the same day: our task groupings had assigned every Opus release to deep reasoning and every Flash release to fast work by inference. GitHub names five models under general-purpose and three under fast, out of 32 — 20 models sit in no group at all. Our inference was this site’s opinion wearing GitHub’s label, which is the same class of error in the opposite direction.

What this does not show

We read the reference page once. We do not know whether it was stale for an hour, a day or a week, and we are not claiming a pattern from one observation.

Nor is this a criticism of the documentation’s accuracy in general. It is a narrow, dated finding about one class of fact — retirement status — and the practical response is to check the changelog for that class rather than to distrust the documentation broadly.

@misc{copilotstack_research-documentation-drift-model-retirements_2026,
  title        = {GitHub's Copilot Docs Contradicted Its Own Changelog on Retirement Day},
  author       = {{The Copilot Stack}},
  year         = {2026},
  note         = {Version 2026-09-04. Maintained research, not peer reviewed},
  howpublished = {\url{https://thecopilotstack.com/research/documentation-drift-model-retirements/}},
  urldate      = {2026-09-04}
}