GitHub Models Officially Shuts Down, Taking Its Inference API, Model Catalog, and BYOK Offline

GitHub Models officially shut down on July 30. This was not a rebranding or a decision to stop accepting new customers: GitHub directly removed the playground, model catalog, hosted inference API, and bring-your-own-key (BYOK) endpoints. Existing customers with active traffic can no longer make calls either. GitHub stopped accepting new customers in June, scheduled brief brownouts on July 16 and 23, and ultimately completed the shutdown as planned.

Models had allowed developers to try models from multiple providers through a roughly unified API and compare prompts and outputs in the GitHub interface. Its removal has particularly significant technical consequences for experimental projects: if an application hard-codes GitHub endpoints, model identifiers, token formats, or response schemas, requests will now fail outright. Even when switching to Microsoft Foundry, teams cannot assume that model names, rate limits, content filtering, tool-calling formats, streaming events, and error codes will be fully compatible.

GitHub officially recommends that applications requiring a model catalog and API move to Microsoft Foundry, while teams that need AI within GitHub workflows should switch to Copilot. However, neither is a one-to-one replacement. Foundry involves Azure projects, regional availability, deployment names, and authentication configuration. Copilot, meanwhile, is designed for development workflows and cannot serve as a general-purpose inference backend for arbitrary applications. Teams that previously used BYOK to access providers through a unified interface must also decide again whether key management and routing should be handled by a cloud platform, a third-party gateway, or an in-house proxy layer.

During migration, teams should first inventory endpoints and model IDs, then use contract tests to verify JSON schemas, function calling, stream termination signals, retry conditions, and content-safety refusals. They must also remeasure latency and token costs and reassess regional data-processing policies, rather than merely replacing the base URL. This shutdown is also a reminder to engineering teams that model aggregation services can disappear entirely. Production systems should isolate provider SDKs behind adapters and retain offline evaluations, switchable credentials and routing, and fallback strategies.