GitHub Models to Shut Down Completely on July 30; Inference Applications Must Migrate to Dedicated Deployment Endpoints

GitHub Models will be fully retired on July 30, 2026. At that point, the playground, model catalog, inference API, and bring-your-own-key (BYOK) interface will all cease operation, including for existing customers that still generate traffic. GitHub scheduled brief outages for July 16 and 23 to expose error-handling issues early in systems that depend on the service. The July 23 brownout is therefore one of the last opportunities for engineering teams to verify that fallback paths, retry limits, and alerts work as intended.

GitHub recommends that projects requiring general-purpose model access move to Microsoft Foundry, while those that only need AI workflows within GitHub should use Copilot. Microsoft Foundry, however, follows a resource–deployment–endpoint model: each deployment is tied to a model name, version, capacity or provisioning method, content-filtering configuration, and rate limits. API calls use the deployment name, so teams cannot assume that existing GitHub Models model identifiers will carry over unchanged. Foundry supports the OpenAI-compatible `/openai/v1/` path and authentication through either an API key or Microsoft Entra ID. When using Entra ID, permissions, token scopes, and service identities must all be incorporated into the deployment process.

The primary migration risks are semantic rather than syntactic. Even when the SDK and request format are compatible, differences in model versions, system-prompt handling, content filtering, tool-call formats, streaming events, context limits, and rate-limiting policies can still alter outputs. Teams should first inventory every GitHub Models endpoint and BYOK dependency, pin replacement model versions, and create a regression suite covering structured outputs, tool calls, long-context requests, and refusal scenarios. They should then run both systems in parallel with a small share of traffic to compare latency, error rates, quality, and cost. Keys should not be copied directly into the new environment; a safer approach is to create new identities, apply least-privilege access, and rotate the old credentials.

Another deadline is also worth noting: Microsoft documentation states that the Azure AI Inference beta SDK will be retired on August 26 and recommends migrating to the stable OpenAI SDK and OpenAI/v1 API. Teams migrating now should avoid integrating an SDK that is already scheduled for retirement, or they may face a second rework within a month. Model and regional availability, quotas, data governance, and pricing also require review, as they are not guaranteed to match GitHub Models.