GitHub Models is retired: your free AI endpoint is gone

As of yesterday, GitHub Models is gone. GitHub's changelog confirms the full retirement on July 30, 2026: the playground, the model catalog, the inference API, and bring-your-own-key. Not deprecated, not read-only — removed, for every customer including the ones with active usage. If a workflow of yours called that endpoint, it started failing yesterday. The GitHub Models retirement is worth five minutes of your attention even if you never used it, because of how the ramp ran.

What actually happened

GitHub telegraphed this properly, which is the part to study. The product closed to new customers on June 16. The full retirement notice landed July 1 with two scheduled brownouts on July 16 and July 23 — deliberate short outages so teams would discover their dependency before the hard cutoff instead of after it. Then July 30, everything off.

The suggested landing spots are Microsoft Foundry for a broad model catalog, or GitHub Copilot if you want models inside GitHub itself. Neither is a drop-in for a free inference API you were hitting from CI.

That's a clean, well-run sunset. It's also a thirty-day window between "this is going away" and "this is gone," for an endpoint that was trivially easy to wire into an Action.

Why a retired model endpoint matters for your business

Free inference bundled into a platform you already pay for is the most seductive integration there is. No new vendor, no new invoice, no procurement conversation. It's also the integration with the least protection — nobody signed anything, so nobody owes you a migration path.

Watch how these things get load-bearing. Someone prototypes a PR summarizer against a free endpoint. It works. Six months later it's classifying support tickets, and the base URL is hardcoded in three workflow files that nobody has opened since. There was never a decision to depend on it.

The fix is boring and takes an afternoon: every model call goes through one adapter in your own code. One module that owns the base URL, the key, and the request shape; everything else calls your function, not the vendor's. Swapping providers becomes a config change and a diff in one file instead of an archaeology project. Same discipline as never scattering raw SQL through a codebase — you already know this pattern, you just have to apply it to the model layer.

Then go grep. github.ai, models.inference, any base URL that isn't yours. Do it before the next sunset notice, not during one.

Sources: GitHub Changelog — GitHub Models is now retired, GitHub Changelog — full retirement notice, GitHub Changelog — closed to new customers.