GitHub Models Dies July 30 — Here's Where to Migrate Your No-Code AI Backend

At 15:59 UTC on July 30, 2026, Microsoft kills GitHub Models, the free AI playground that let developers test 28 models directly inside GitHub. If your no-code app or workflow hits the models.inference.ai.azure.com endpoint for inference, it stops working at that exact minute. Not degrades. Not rate-limits. Stops.

This one's been coming in stages. June 16: closed to new customers. July 1: full retirement announced, with two scheduled brownouts already behind us (July 16 and July 23. If you missed those interruptions, you might already be in trouble). And now we're a week and change from the final switch-off.

I've been saying for a while that free AI sandboxes are scaffolding, not foundation. GitHub Models is the latest proof point. Here's what you need to do.

What was GitHub Models, anyway?

Launched in 2024, GitHub Models was a properly useful thing: a model playground sitting right next to your repos. You could test prompts against Llama 4, Mistral Large 3, Cohere Command R7, Microsoft's own Phi family, and OpenAI's GPT-4o. No API key wrangling, no separate billing relationship. Just a GitHub token and you were off.

It was popular with no-code builders for exactly that reason. Prototyping an AI feature? Point your Bubble workflow or your Make.com automation at the GitHub Models endpoint and test with a dozen different models for free. The rate limits were generous enough for development. The friction was near zero.

In mid-2025, Microsoft added pay-as-you-go billing and bring-your-own-key support, which signalled the direction of travel. But the free tier remained, and that's what most people were using. The whole thing ran on Azure inference underneath, which makes what happened next entirely predictable.

Who actually needs to act?

Three groups.

First: anyone who hard-coded the GitHub Models API into a no-code automation. If you've got a Make.com scenario, a Zapier zap, or a Bubble plugin action pointing at models.inference.ai.azure.com/chat/completions, that call dies on July 30. This isn't a deprecation with a grace period. It's a hard cutoff.

Second: teams running AI calls inside GitHub Actions workflows. If your CI/CD pipeline calls GitHub Models for PR summaries, issue triage, or code review, find those calls now. The brownouts on July 16 and 23 were supposed to surface these dependencies, but if you weren't watching your logs closely, you might have missed them.

Third: anyone who was about to start prototyping. You're already locked out if your org never used GitHub Models before June 16. Don't waste a sprint designing around something you can't access.

Where do you go now?

The official off-ramp is Azure AI Foundry. Microsoft points everyone there, and it makes sense if you're already in the Azure/Entra ecosystem. Pricing for models like GPT-4o runs roughly $2.50 per million input tokens and $10 per million output tokens. Not free, but it comes with proper SLAs and enterprise billing. The trade, of course, is that you're trading one Microsoft dependency for another, deeper one.

I'd suggest thinking about this differently, though. The real lesson here isn't find another free tier. It's that any AI endpoint you can't swap in an afternoon owns you.

OpenRouter is the portability play. One API key, one OpenAI-compatible endpoint, hundreds of models behind it. If a provider pulls a model or jacks the price, you change a string in your config, not your entire integration. For no-code builders, this is the closest thing to future-proofing you can get without self-hosting.

Direct provider APIs (Anthropic, OpenAI, Google) make sense when you've settled on one model and want the best latency and newest features. Less flexible, more control.

And if your usage is light or your data shouldn't leave your machine, run models locally with Ollama or LM Studio. Open-weight models like Llama 4 and Mistral Large 3 have capable quantised versions that run on a decent laptop. Zero per-token cost, zero vendor risk.

Here's the thing, though: the smartest move isn't picking the right replacement. It's not wiring yourself to any single provider in the first place.

Most of these services speak the OpenAI-compatible chat completions format. That means you can put a thin abstraction in front of your AI calls. An environment variable for the base URL, another for the model name. Swapping providers becomes a config change instead of a rebuild. Some no-code platforms handle this natively. If yours doesn't, it's worth finding one that does.

The pattern keeps repeating

This isn't just a GitHub story. Free AI playgrounds keep getting retired the moment the company behind them figures out how to charge for the real thing. Google's done it. OpenAI's done it. Microsoft just did it. The free tier is always a loss leader for something, and when that something matures, the free tier evaporates.

The no-code platforms that survive these shifts are the ones built on model-agnostic architecture. The AI backend is a configurable layer, not baked into the platform's DNA. If you're on a tool that ties you to a specific provider and that provider changes the deal, your entire app breaks. If your platform lets you point at whatever endpoint you want, the GitHub Models shutdown is a five-minute config change and a cup of tea.

The takeaway: grep your codebase and your automations for models.inference.ai.azure.com today. Pick a replacement based on portability, not convenience. And if your no-code platform can't swap AI backends without a rebuild, start asking questions about what happens when (not if) the next provider pulls the same move.