GitHub Models Shuts Down July 30 - Here's Your Two-Week Migration Window

Deadline stories are rare in this space - most tool news is "something new exists, evaluate at your leisure." This one has a date, and the date is nineteen days away.

GitHub announced on July 1 that GitHub Models "will be fully retired on July 30, 2026" (GitHub Changelog). Not deprecated, not new-signups-closed: retired. The playground, the model catalog, the inference API, and the bring-your-own-key endpoints all stop working, and the UI gets removed. GitHub is unusually blunt about scope: "This affects all customers. Unlike the first step in June, this change applies to everyone, including existing customers with active usage."

If you never touched GitHub Models, you can stop reading - nothing here changes your stack. If you did, here's the triage.

First, the calendar

Three dates matter, and two of them arrive before the shutdown:

• July 16 - first scheduled brownout. GitHub Models requests "will temporarily return errors before service is restored."
• July 23 - second brownout, same deal.
• July 30 - permanent shutdown of playground, catalog, inference API, and BYOK.

The brownouts are the useful part, oddly. They're a free chaos-engineering exercise: whatever breaks in your systems on July 16 is exactly what breaks forever on July 30, except you get to fix it with two weeks of runway. If you have anything in production that touches a GitHub Models endpoint, watch your error dashboards on the 16th rather than discovering the dependency on the 30th.

What GitHub Models actually was, and why people used it

For those who adopted it, the appeal was friction removal: a free-to-cheap way to prototype against many different models - compare a prompt across providers, wire a quick inference call into a GitHub Action, test with your own API keys through the BYOK path - without leaving GitHub or standing up separate provider accounts. It became a quiet dependency in a lot of hobby projects, CI experiments, and internal tools precisely because it was the lowest-effort option.

That's what makes the shutdown notice worth taking seriously. Low-friction services accumulate forgotten dependencies. The call you wired into a workflow file eight months ago and never thought about again is the one that starts failing.

Where to move, by use case

GitHub's official guidance names two destinations, and they map to different users:

Microsoft Foundry is the recommendation if what you valued was the multi-model catalog - one place to reach many models. It "offers a broad model catalog" and is the closest like-for-like replacement for the catalog-plus-inference-API pattern. The trade-off is that you're now managing an Azure-side service rather than something bundled into your GitHub account.

GitHub Copilot is the recommendation if your usage was really about AI inside GitHub workflows. Copilot "gives you access to a range of models" within the GitHub surface. Two honest caveats: it's a different shape of product (assistant seats, not a raw inference API), and as of June 1 it meters usage through AI Credits with no free fallback - so "just use Copilot instead" is not a zero-cost migration for anything with volume.

Direct provider APIs are the unlisted third option, and for a lot of prototyping use they're the right one. Model access has never been cheaper to buy directly: GPT-5.6's Luna tier runs $1/$6 per million tokens, Claude Sonnet 5 is $2/$10 at intro pricing, and Grok 4.5 sits at $2/$6. If you were using GitHub Models to avoid the overhead of provider accounts, that overhead is mostly what you're buying back - and in exchange you get pricing and rate limits that are contractually yours rather than a platform feature that can be retired.

The pattern behind the shutdown

Read alongside the Copilot billing change, the direction is hard to miss: GitHub is consolidating every AI touchpoint into Copilot, its credit-metered flagship, and shedding the adjacent free surfaces. That's rational product strategy, but it carries a lesson for tool choosers that outlasts this particular deprecation - platform-bundled AI conveniences are features, not services, and features get rationalized when strategy shifts. For anything load-bearing, prefer the dependency you pay for directly over the one that came free with a platform.

Your checklist, in order

1. Grep your repos, workflow files, and internal tools for GitHub Models endpoints and API references. Do this before July 16, then let the brownout confirm you found everything.
2. Classify each hit: dead experiment (delete), prototype (move to a direct provider API), or production dependency (decide Foundry vs direct provider this week).
3. If any BYOK usage exists, note that those endpoints die on the same date - your own keys don't keep the service alive.
4. Re-run the check after the July 23 brownout. Anything that errored is your final punch list for the last week.

Nineteen days is plenty if you start now. It's very little if you find out on the 30th.