GitHub Models reaches full retirement on July 30, 2026. The shutdown removes the built-in playground, model catalog, inference API, and bring-your-own-key endpoints for every customer, including teams with active usage.
This is not a cosmetic change or a renamed interface. Applications calling GitHub Models must move to another inference provider, and evaluation workflows built around its catalog and playground need a new home.
What disappeared
GitHub first closed Models to new customers in June. Its July 1 retirement notice extended the change to existing users and scheduled brownouts on July 16 and July 23 so teams could see how their systems behaved when requests failed.
The final scope is broad: the user interface, hosted model discovery, direct inference, and BYOK access all leave together. That makes the migration larger than changing one API base URL. Teams may need to replace authentication, model identifiers, request formats, rate-limit handling, evaluation tooling, and cost controls.
Two destinations for two different jobs
GitHub points developers toward Microsoft Foundry when a product needs direct access to a broad model catalog. Foundry is the infrastructure path: teams choose and integrate models into their own applications.
For AI-assisted work inside GitHub, the company recommends GitHub Copilot. That is the workflow path: coding assistance and agentic development remain attached to repositories, issues, pull requests, and development environments.
The distinction matters. A model API is a product dependency; an AI coding assistant is a development tool. Teams using GitHub Models for production inference cannot treat Copilot as a drop-in replacement.
What teams should migrate now
Start with an inventory of every Models endpoint, token, model name, evaluation script, and playground-derived prompt in active projects. Then map each dependency to the replacement provider and test the operational differences rather than only checking whether one successful response comes back.
Migration tests should cover output quality, latency, streaming behavior, safety filtering, retry logic, quotas, regional availability, observability, and cost. BYOK users also need to confirm how credentials and provider-specific policies are handled in the new architecture.
Why it matters
GitHub Models offered developers an unusually convenient place to compare and integrate multiple AI models without leaving the developer platform. Its closure removes that neutral experimentation layer from GitHub.
GitHub has not stated that consolidation is the motive. The product direction nevertheless produces a clearer division across Microsoft’s stack: general-purpose model infrastructure in Foundry, with GitHub concentrating its own AI experience around Copilot.
The product lesson
Abstraction layers make model experimentation fast, but they can also hide switching costs until a platform changes direction. Production teams should keep provider boundaries explicit: isolate model calls, own evaluation datasets, preserve portable prompts, and make authentication and routing configurable.
That architecture will not make every migration painless. It does turn a platform shutdown from an emergency rewrite into a controlled provider change.