GitHub Just Closed Its Free AI On-Ramps. Models Is Retired to New Users and Copilot Sign-Ups Got Paused. If You Prototype on Free Vendor Inference, This Is the Lesson.

On June 16, GitHub announced that GitHub Models is no longer available to new customers. If your org or enterprise hadn't used it before, you won't see it and can't start. Existing users keep working for now, and GitHub points new projects to Azure AI Foundry instead. Around the same stretch, GitHub has been pausing and reworking new Copilot Individual sign-ups as AI coding drives up compute demand.

Put those together and the pattern is clear: the free, low-friction AI on-ramps inside GitHub are closing, and the path forward runs through a billing relationship with Azure. If you've been prototyping on free vendor inference, this is the reminder that a free playground is a demo, not infrastructure.

What's actually happening

Two moves, same direction.

GitHub Models (the playground and API that let you try a catalog of models inside GitHub) is being retired, starting with new customers. Per GitHub's own changelog, existing customers with active usage aren't affected today, but new orgs and enterprises are locked out on both free and paid plans, and GitHub says it'll share full retirement timelines later. The recommended replacement is Azure AI Foundry, which is Microsoft consolidating AI tooling under Azure rather than running a parallel offering inside GitHub.

Separately, GitHub has throttled new Copilot Individual sign-ups, with reporting tying it to compute demand outrunning supply. Different product, same underlying reality: serving AI to everyone for free or near-free is expensive, and when the bill gets big enough, the open door becomes a managed one.

The thread connecting them is that the cheap, frictionless way in is the first thing to get pulled when a feature gets popular enough to cost real money.

Why a free playground is a demo, not infrastructure

There's a specific trap here, and it's one I've walked into. You build a prototype against a free model endpoint because it's right there, zero setup, no credit card. It works. You keep building. Six weeks later the prototype is load-bearing, real users depend on it, and you never went back to put it on a foundation you actually control. Then the free endpoint changes its terms, caps new access, or gets retired, and you discover your product was sitting on a vendor's marketing budget the whole time.

Free vendor inference exists to get you in the door. That's its job, and it's an honest job: try-before-you-buy is fine. The mistake is mistaking the on-ramp for the road. The provider will keep the door open exactly as long as it's cheap to, and "popular" and "cheap" stop being compatible at scale. GitHub Models getting retired to new users and Copilot sign-ups getting paused are the same lesson told twice in one month.

This isn't a betrayal and I want to be clear about that. Microsoft isn't doing anything wrong by consolidating inference under Azure and metering it. Inference costs money, free tiers are subsidies, and subsidies end. The cost just lands on whoever built a product on top of the subsidy without noticing.

What to standardize on instead

The fix is to own the parts of your inference path that you can't afford to have pulled out from under you.

Own your API key and your billing relationship for anything in production. A direct, paid relationship with a model provider doesn't get "retired to new customers" out from under you the way a free convenience layer does. You're a customer with a contract, not a guest on a free tier.

Route model calls through a gateway or proxy you control, not hard-coded to one vendor's convenience endpoint. If your code calls an abstraction you own, swapping the provider behind it is a config change. If it calls a specific free playground API directly, swapping it is a rewrite under deadline pressure after the thing already broke.

Keep a prototyping path that survives a vendor pulling its free tier. It's fine to prototype fast on whatever's frictionless, just don't let "frictionless" and "in production" be the same dependency. The moment a prototype starts mattering, move it onto something you'd be comfortable paying for, before you're forced to.

The honest counter-take

I'll argue against my own urgency for a second, because not everything needs to be migrated this weekend.

If you're an existing GitHub Models user, nothing breaks today. You have a window, and you should use it calmly rather than panic-porting everything to Azure AI Foundry because a changelog spooked you. Azure AI Foundry is also a perfectly reasonable destination if your stack already lives in the Microsoft world; consolidation isn't automatically bad, and one well-supported platform can beat three half-maintained ones. The risk isn't "Microsoft is the wrong choice." The risk is ending up somewhere by default, because a free tier closed and you took the path of least resistance, instead of choosing your inference foundation on purpose.

The point isn't which vendor. It's that you picked, you own the key, and you can leave.

What I'd actually do

Audit your stack for anything in production that calls a free or convenience model endpoint you don't have a paid relationship with. Those are your exposed points: a vendor can close them the way GitHub just closed Models to new users. Move each one behind a gateway you control, on a key you pay for, with a second provider you've actually tested as the fallback.

Then keep prototyping on whatever's frictionless. Just promote it to a foundation you own the day it starts to matter, not the day it breaks.

Author

Lukas

@lukcombinator

Sources

• GitHub Models is no longer available to new customers (GitHub Changelog)
• GitHub Retires Its Free AI Model Playground: What Developers Need to Know (DevOps.com)
• GitHub pauses Copilot sign-ups as AI coding drives up compute demand (The New Stack)
• Changes to GitHub Copilot Individual plans (The GitHub Blog)