GitHub Models Is Retired: What to Use for Model Access Now

GitHub Models is officially retired. As of July 30, 2026, the playground, model catalog, inference API, and bring your own key (BYOK) are no longer available to any customer, including existing customers with active usage. The change was confirmed in the GitHub Changelog on July 30, which closes out a retirement process that started in June.

If you built a prototype, a CI workflow, or an internal tool on top of GitHub Models' token-based inference API, this is a migration event, not a background note. Here is exactly what happened, what it means, and where the model access went.

What actually changed#

GitHub Models launched in 2025 as the lowest-friction way to call frontier models: a playground plus an inference API authenticated with your GitHub token, no separate account, no credit card, no per-provider API key. It became a common on-ramp for experiments in Codespaces and GitHub Actions, and BYOK let teams pin their own provider keys on top of the same interface.

The shutdown happened in three steps, all documented in the official changelog:

1. June 16, 2026: GitHub Models closed to new customers. Organizations and enterprises without existing usage lost access on both free and paid plans.
2. July 1, 2026: Full retirement announced for July 30, including a warning about scheduled brownouts (temporary service interruptions) on July 16 and July 23.
3. July 30, 2026: The service is gone. The announcement states the playground, model catalog, inference API, and BYOK "are no longer available to any customer, including existing customers with active usage."

The July 1 notice is notable for how it treats the final step: it applies to everyone. The June step grandfathers existing users; the July step does not.

Why this matters to developers#

The retirement removes one of the few zero-setup ways to evaluate models and ship a working AI feature in a day. Three kinds of projects are affected:

• Prototypes and hackathon builds that called the inference API with a GitHub token. Those endpoints return errors now, and the playground UI is gone.
• CI pipelines that used GitHub Models for smoke-testing prompts or evaluating outputs in Actions. The authentication model was the whole point: no separate key to rotate or bill.
• BYOK configurations, which funneled provider keys through GitHub's endpoint. Those integrations are dead even though the underlying provider accounts still exist.

For most teams the migration is small: swap the endpoint and credentials, keep the prompt logic. But the free tier was a real distribution channel for model evaluation, and its absence makes direct provider accounts or a gateway the default starting point again.

Where model access lives now#

GitHub's own guidance points in two directions, both documented in the July 30 changelog:

• Microsoft Foundry (ai.azure.com) offers a broad model catalog and is the named replacement for new and existing projects. We have a full walkthrough of running Claude through Foundry with a dev-friendly setup in our Foundry guide.
• GitHub Copilot remains the path for model-powered workflows that live inside GitHub, and it supports a range of models behind the same subscription model.

Beyond GitHub's recommendations, the realistic options are the same ones that existed before GitHub Models made them optional:

• Direct provider APIs from OpenAI, Anthropic, Google, and xAI. This is the most flexible route and pairs with the API access patterns we cover in our GPT-5.6 developer guide.
• Gateway and routing layers, which abstract multiple providers behind one key and let you switch models without touching application code. If your project consumed GitHub Models as a neutral endpoint, a router is the closest mental replacement.
• Open-weight models with cheap or self-hosted inference. For evaluation workloads and batch jobs, models in this category keep the economics of the old free tier while removing the vendor dependency entirely.

The pattern behind this shutdown#

This is not an isolated incident. Model APIs are consolidating around fewer, stricter on-ramps: free tiers shrink, endpoints get cut, and access moves behind subscriptions or enterprise agreements. We covered the parallel story when OpenAI retired older GPT model versions and forced migrations. The same math applies here: if your tooling calls a model endpoint you do not control, budget for the endpoint's lifecycle, and keep the model layer thin so the swap is a configuration change rather than a rewrite.

What to do today#

If you were an active GitHub Models user, three moves cover most cases:

1. Inventory the call sites. Grep for the GitHub Models endpoint and any token-based auth in your repos. The brownout dates (July 16 and 23) were the dry run; the errors you see now are the real thing.
2. Pick a replacement per workload. Prototypes can move to Foundry or a direct provider key. CI evaluation jobs are often the best fit for a router or open-weight model to keep cost near zero.
3. Rewrite credentials, not logic. The OpenAI-compatible request shape carried by most endpoints means your prompt engineering and parsing code should port with minimal changes.

The takeaway: GitHub Models was a convenient door, not a platform. The projects that treated it as a configurable endpoint migrate in an afternoon. The ones that built around it as a platform now have a small rewrite in front of them, and this is the second such migration in a year, so plan the model layer to outlive any single provider.

Continue Reading#

• Claude on Microsoft Foundry: Developer Guide
• Migrating Off Retired GPT Models: The Playbook
• AI Model Routing as an Orchestration Layer
• GPT-5.6 Developer Guide: Access and API Patterns
• DeepSeek Chat to v4 Migration Guide

Sources#

• GitHub Models is now retired (GitHub Changelog, July 30, 2026)
• GitHub Models is being fully retired on July 30, 2026 (GitHub Changelog, July 1, 2026)
• GitHub Models is no longer available to new customers (GitHub Changelog, June 16, 2026)
• GitHub Models documentation
• Microsoft Foundry