# GitHub Models Got 29 Days. Plan Your AI Stack Accordingly > Source: > Published: 2026-08-03 01:08:16+00:00 [AI](https://sourcefeed.dev/c/ai)Article # GitHub Models Got 29 Days. Plan Your AI Stack Accordingly Microsoft retired its free inference API with under a month's hard notice, paying customers included. That's the real lesson. [Mariana Souza](https://sourcefeed.dev/u/mariana_souza) GitHub Models is gone. On July 30, GitHub [switched off](https://github.blog/changelog/2026-07-30-github-models-is-now-retired/) the playground, the model catalog, the inference API, and the bring-your-own-key endpoints — for everyone, including existing customers with active usage and active bills. The service that launched in August 2024 promising to put AI models "in the hands of 100 million developers" didn't make it to its second birthday. The shutdown itself isn't the scandal. Products die; loss leaders die faster. What should actually change how you build is the timeline. GitHub closed the service to new customers on June 16, announced full retirement on July 1, ran deliberate brownouts on July 16 and 23 to smoke out hidden dependencies, and pulled the plug on July 30. If you were a paying customer — and GitHub had been selling paid tiers beyond the free limits since mid-2025 — you got twenty-nine days of hard notice to migrate a production dependency. That's not a deprecation policy. That's an eviction. ## The freest API in the industry It's worth remembering why GitHub Models mattered, because nothing has cleanly replaced it. You took the GitHub personal access token you already had, pointed an OpenAI-compatible client at GitHub's endpoint, and you were calling GPT-4o, Llama, DeepSeek-R1, or Phi — no credit card, no Azure subscription, no separate signup. Rate limits were tight (dozens of requests a day on the top-tier models), but for prototyping, coursework, hackathons, and CI-driven experiments, it was the lowest-friction model access anywhere. It even wired into Actions workflows, so a repo could call inference with its built-in token. That zero-friction design is exactly why the shutdown stings more than the numbers suggest. GitHub Models wasn't in anyone's architecture diagram. It was in scripts, notebooks, classroom starter repos, and that one workflow someone set up in 2025 and forgot about — the kind of dependency the July brownouts were explicitly designed to make fail loudly, because nobody could find it any other way. ## Why Microsoft killed it GitHub's retirement notice offers no reason, but you don't need one spelled out. Look at the two migration paths it points to: [Microsoft Foundry](https://azure.microsoft.com/en-us/products/ai-foundry) for raw model access, and [GitHub Copilot](https://github.com/features/copilot) for AI workflows inside GitHub. That's the whole story. At Ignite last November, Microsoft rebranded Azure AI Foundry to Microsoft Foundry and positioned it as *the* model platform, a peer to Microsoft 365. Meanwhile GitHub itself was pulled closer into Microsoft's CoreAI organization after Thomas Dohmke stepped down as CEO last August. In that org chart there is no room for a second, mostly-free inference API that competes with the flagship's funnel and monetizes worse than Copilot. Free model inference was always a marketing expense — a way to get developers touching Azure-hosted models. Marketing budgets get reallocated. The mistake is reading a strategic promotion as infrastructure. ## What migration actually looks like If you used GitHub Models through its OpenAI-compatible endpoint, you're in decent shape: migration is a new base URL, a new key, and a model-ID mapping. That seam — the OpenAI wire format as lingua franca — is the one piece of genuinely good news in this story, and it's why the practical damage is measured in hours, not weeks. But none of the destinations is a drop-in replacement for what made GitHub Models useful: **Microsoft Foundry** has the catalog, but it wants an Azure subscription, a resource group, and deployment decisions before you send your first token. That's fine for teams already on Azure; it's a wall for the student who had a PAT and nothing else.**Copilot** answers a different question. It's a product you use, not an API you build on — pointing API users there is a category error dressed up as a migration path.is the closest spiritual successor: one key, one OpenAI-compatible endpoint, hundreds of models, and — crucially — it's an aggregator whose entire business is portability, so its incentives point away from lock-in rather than toward it.[OpenRouter](https://openrouter.ai)**Local models via** now cover the "free experimentation" tier better than any hosted freebie, with no rate limits and no retirement risk. For prototyping, a local Llama or Qwen behind the same OpenAI-compatible interface is the durable answer.[Ollama](https://ollama.com) ## Deprecation policy is a feature — score it The dev.to post that flagged this story draws the standard lesson: hide providers behind a narrow adapter so a shutdown is a config change, not a rewrite. That's correct and twenty years old; it applied to payment processors and map APIs long before LLMs. The newer, sharper lesson is about *vendor selection*, not code structure. AI inference is currently the most volatile dependency category in software. Models get deprecated on twelve-month cycles, pricing gets restructured quarterly, and — as of July 30 — a Microsoft-owned platform has demonstrated it will give paying customers under a month to leave. So when you evaluate a model provider, the deprecation track record deserves the same weight as latency and price. OpenAI and Anthropic publish model deprecation schedules with long runways; Google has retired whole APIs (PaLM, take a bow) with more notice than GitHub gave here. GitHub just set the floor for how fast a hyperscaler subsidiary will move when strategy shifts, and every "free AI credits for developers" program you see today should be read against it. The practical checklist is short. Keep provider calls behind one adapter and model IDs in config, yes. But also: run your evals against at least two providers so switching is a decision, not a research project; treat any free tier as having a 12-month half-life; and prefer vendors whose business *is* inference over vendors for whom it's a growth hack. GitHub Models was a great growth hack. Nobody should have been surprised — and after this, nobody gets to be surprised twice. ## Sources & further reading - [GitHub Models is being fully retired on July 30, 2026](https://github.blog/changelog/2026-07-01-github-models-is-being-fully-retired-on-july-30-2026/)— github.blog - [GitHub Models is now retired](https://github.blog/changelog/2026-07-30-github-models-is-now-retired/)— github.blog - [GitHub Models Shut Down: What Beginners Should Learn About AI Vendor Lock-In](https://dev.to/marcusykim/github-models-shut-down-what-beginners-should-learn-about-ai-vendor-lock-in-3d3p)— dev.to - [Introducing GitHub Models: A new generation of AI engineers building on GitHub](https://github.blog/news-insights/product-news/introducing-github-models/)— github.blog - [GitHub Retires Its Free AI Model Playground - What Developers Need to Know](https://devops.com/github-retires-its-free-ai-model-playground-what-developers-need-to-know/)— devops.com [Mariana Souza](https://sourcefeed.dev/u/mariana_souza)· Senior Editor Mariana covers the fast-moving world of machine learning and generative AI, with a particular focus on how these technologies are reshaping development workflows. When she isn't stress-testing the latest foundation models, she's usually at a local hackathon. ## Discussion 0 No comments yet Be the first to weigh in.