GitHub is turning off GitHub Models on July 30, 2026. Not narrowing it, not moving it behind a paywall — retiring it fully. The playground, the model catalog, the inference API, and bring-your-own-key all stop working for every customer, and two brownouts on July 16 and July 23 will break your integration on purpose before the real cutoff, so a silent prod failure surfaces while you can still fix it.
If you wired a prototype, a CI job, or a demo to https://models.github.ai/inference with a GitHub PAT, here is the whole decision in one screen — then the detail per path.
The one-screen answer#
| Destination | Effort | Per-token cost | Pick it when |
|---|---|---|---|
| Azure AI Foundry (GitHub's pick) | Heavy | Metered | You're already shipping on Azure |
| OpenRouter | Light | Metered + small markup | You want "many models, one endpoint" back |
| Direct provider key | Light | Metered, list price | You've settled on one model |
| Local via Ollama | Medium | Free per token | You want free + private dev/CI |
The good news underneath the panic: GitHub Models implemented the OpenAI Chat Completions format, and so does every path above. That makes three of the four a base-URL-and-key change, not a rewrite. We wrote the exact endpoint swap up separately — migrate off GitHub Models in 15 minutes — so this piece is the which, not the how.
There is no soft landing#
Two facts make this different from a normal deprecation, and both argue for moving now rather than in the last week of July.
First, there is no grandfathering and no paid tier to fall into. New customers were already cut off in June; on July 30 the endpoint stops answering for everyone else. You cannot buy your way to a few more months.
The brownouts on July 16 and 23 aren't a courtesy — they're a fire drill GitHub is running for you. If your integration is going to break, it breaks on those days, on GitHub's schedule, instead of at 2 a.m. on the 30th on yours.
Second, GitHub Models' whole appeal was zero setup: a token you already had, an endpoint that already spoke OpenAI, no billing. Every replacement gives back some of that friction. The question is which friction you can most afford.
Azure AI Foundry — the official path, and the heaviest#
GitHub points you at Azure AI Foundry, and for the right team it's the best answer: the broadest model catalog, enterprise controls, and a straight line to production on Microsoft's cloud. The catch is that Foundry needs an Azure subscription, billing, and resource provisioning — precisely the setup GitHub Models existed to skip.
Take this path if you were going to build on Azure regardless, or if you need Foundry's governance and catalog breadth. Don't take it reflexively just because it's the recommended button. For a solo founder who wanted a free endpoint to poke at models, Foundry is a lot of platform to stand up.
OpenRouter — the closest thing to what you're losing#
What GitHub Models really gave you was many models behind one URL and one key. OpenRouter is the most direct replacement for that experience: a single OpenAI-compatible endpoint fronting hundreds of models, so you keep the "try three models by changing one string" workflow. You pay per token with a small routing markup, and you swap providers without re-plumbing. If you're not sure which model you'll land on, this buys you time to decide without another migration. (If you're weighing it against a self-hosted router, we compared OpenRouter vs LiteLLM and the three big gateways directly.)
A direct provider key — the lowest-latency path once you've chosen#
If your prototyping already told you the answer — you want GPT-5.6, or Claude, or a cheap open-weight coder — go straight to that vendor. OpenAI is natively compatible; Anthropic exposes an OpenAI-compatible endpoint; Moonshot's Kimi K2.7 Code is OpenAI-compatible and cheap for agentic coding. One key, one base URL, no middle layer taking a cut or adding a hop. This is the play when you know your model and want the shortest path to it.
Local via Ollama — if "free" was the whole point#
Plenty of teams used GitHub Models simply because it was free and private. Keep both by running the model yourself with Ollama, which just raised a $65M Series B on exactly this thesis. Per-token cost drops to zero for development, CI, and anything that can't leave your environment. The trade is hardware: big models want a real GPU, so the practical shape is a small local model for dev and a hosted provider for production traffic.
The move that outlasts this one#
Whatever you pick, don't hard-code the endpoint again. Put the base URL and key behind one environment variable and route every call through a single thin client. Because all of these providers speak the OpenAI format, that indirection is nearly free — and it turns the next deprecation from a scramble into a one-line config change. GitHub Models won't be the last free endpoint to disappear; the teams that shrug it off are the ones who already made their model provider a variable, not a constant.
