GitHub Models was fully retired on July 30, 2026. The shutdown affects every customer and removes the GitHub Models playground, model catalog, inference application programming interface (API), and bring-your-own-key endpoints. Developers with applications, evaluations, or demos that call GitHub Models must move those workloads to another model provider; changing only a token will not restore a retired GitHub endpoint. GitHub recommends Microsoft Foundry for applications that still need model inference and GitHub Copilot for AI workflows inside GitHub. The practical priority is to inventory every dependency, select and deploy replacement models, update endpoint and authentication settings, then compare outputs and operational controls before sending production traffic to the replacement.
Key takeaways
- GitHub says the entire GitHub Models service is retired as of July 30, 2026, including its user interface and inference endpoints.
- Microsoft Foundry is the closest officially recommended path for teams that need a multi-model inference service, but it requires an Azure subscription and explicit model deployments.
- GitHub Copilot is a coding and workflow product, not a drop-in replacement for an application that called the GitHub Models API.
- A safe migration covers prompts, model identifiers, authentication, endpoint URLs, rate limits, safety controls, evaluations, and observability—not just SDK syntax.
- Model availability, regional support, throughput, and billing depend on the selected Foundry deployment. Confirm them in the portal before committing.
What GitHub retired
GitHub announced on July 1 that GitHub Models would be fully retired on July 30, 2026. Its notice explicitly includes the playground, model catalog, inference API, and bring-your-own-key (BYOK) endpoints. The retirement applies to existing customers with active usage, not only to new sign-ups. GitHub also scheduled temporary brownouts on July 16 and July 23 as advance warnings. GitHub’s retirement notice is the controlling source for the scope and date.
This means four common workflows need attention:
- Interactive prompt experiments: saved practices that depended on the GitHub playground need another experimentation interface.
- Application inference: calls to GitHub-hosted Models endpoints need a new host, credentials, and usually a deployment-specific model name.
- Catalog discovery: scripts or internal tools that queried the GitHub Models catalog can no longer treat it as a live source.
- BYOK routing: supplying a provider key through GitHub Models does not preserve access because GitHub is removing the routing endpoint itself.
Repository files such as prompts, fixtures, and evaluation datasets are ordinary Git content and are not inherently deleted by the product retirement. However, any automation that invokes the retired service will fail until it is redirected.
Which replacement fits each workload?
GitHub names two alternatives, but they solve different jobs. Its retirement notice points model-access applications toward Microsoft Foundry and GitHub-native AI workflows toward GitHub Copilot.
| Existing use | Likely path | Why | Main caveat |
|---|---|---|---|
| Application calls a model API | Microsoft Foundry Models | Provides deployable model endpoints and a model catalog | Requires Azure setup, deployment selection, and billing review |
| Team compares prompts and model outputs | Foundry playground plus your own evaluation harness | Preserves interactive exploration and repeatable tests | Existing GitHub evaluation artifacts may need manual adaptation |
| Developer wants coding assistance in GitHub | GitHub Copilot | Designed for code and agent workflows within GitHub | Not a general-purpose inference backend |
| Small provider-specific integration | Direct model-provider API | Fewer platform layers and provider-native features | Less portability across model families |
| Regulated or centrally governed workload | Existing approved cloud AI platform | May fit established identity, logging, and procurement controls | Feature and model parity must be checked |
The direct-provider and existing-cloud options are editorial alternatives, not GitHub recommendations. Choose them only when they reduce operational complexity or align with identity, networking, billing, and regional controls already in place. Compare those controls against live vendor documentation before a procurement decision.
How Microsoft Foundry differs from GitHub Models
Microsoft’s official GitHub Models upgrade guide says Foundry requires an Azure subscription with a valid payment method. Partner and community models can also require permission to subscribe to Azure Marketplace offers. Unlike GitHub Models, where models were preconfigured for experimentation, Foundry requires teams to deploy the models they intend to call.
That change introduces useful control but also more responsibility. A Foundry deployment combines a model and version with capacity or provisioning choices, content-filter settings, and rate-limit configuration. Microsoft says one Foundry resource can contain multiple deployments, each addressed by its deployment name. Its endpoint documentation also supports Microsoft Entra ID authentication, allowing teams to avoid long-lived API keys after the required permissions are configured.
Microsoft’s upgrade documentation says compatible application code may require no other changes after the key and endpoint are replaced. Treat that as a best-case path, not a verification result. The hostname, credential, deployment name, available parameters, response behavior, filters, quotas, and billing context can all differ. Production owners should prove compatibility with their own requests and acceptance tests.
The destination architecture must also cover gateways, rate limits, request and audit logging, retries, timeouts, identity, and regional routing. Model selection should be driven by representative task results and operational constraints, not by treating the migration target as a permanent default.
A practical migration plan
1. Find every GitHub Models dependency
Search source, infrastructure definitions, continuous integration secrets, notebooks, documentation, and monitoring configuration for GitHub Models hostnames, SDK clients, model IDs, and token variable names. Check scheduled jobs and dormant branches as well as the primary application.
Record each workload’s owner, environment, model, endpoint, authentication method, prompt or request format, expected output, traffic level, latency tolerance, and data classification. This inventory prevents a successful migration of the main service from hiding failures in evaluation jobs or internal demos.
2. Preserve the behavior you need to compare
Before rewriting the integration, retain representative request fixtures and expected properties of acceptable responses. Do not store sensitive production prompts in an unsecured test set. Capture structural requirements—valid JSON, required fields, citation presence, refusal behavior, tool calls, or maximum response length—rather than expecting word-for-word output.
If the workload uses retrieval-augmented generation (RAG), test retrieval and generation separately. An endpoint migration should not silently change chunking, retrieval, grounding, and generation at the same time because that makes failures difficult to attribute.
3. Select a target model and deployment
In Microsoft Foundry, verify that the intended model is available in an acceptable Azure region and deployment type. Confirm that the model supports the input modalities, output format, context needs, tools, and safety settings used by the application. Do not assume a familiar model name guarantees the same version or configuration.
Name the deployment independently from the underlying model when that will make future upgrades safer—for example, an application-facing alias that can move only after validation. Microsoft states that requests use the deployment name in the model parameter.
4. Update endpoint and authentication
Microsoft’s Foundry endpoint documentation describes a generally available OpenAI-compatible /openai/v1/ route and deployment-based endpoints. It also warns that the Azure AI Inference beta software development kit (SDK) is deprecated. For a fresh migration, avoid creating a second near-term migration by adopting the current stable API and SDK path shown in the documentation.
Keep credentials in a secret manager, not in the repository. Where the operating environment supports it, consider Microsoft Entra ID with narrowly scoped identities. If API keys are required, rotate them, separate environments, and define an owner and expiry process.
5. Run compatibility and safety checks
Send a bounded evaluation set through both the recorded baseline and the replacement when a baseline is still available. If it is not, compare the new result against explicit acceptance criteria. Measure schema validity, task success, refusal behavior, groundedness, token usage, latency, and error handling.
Recheck content filters and blocked-content responses. A stricter or looser default can change application behavior even when ordinary prompts appear equivalent. Validate streaming, retries, timeouts, tool calls, embeddings, and rate-limit handling separately because success on one chat request proves little about those paths.
Record the fixtures, acceptance criteria, failure categories, latency distribution, token use, manual correction effort, and cost so the same evaluation can be rerun after the migration.
6. Cut over with rollback boundaries
Route a small, observable portion of eligible traffic to the new deployment if the architecture permits it. Define failure thresholds before the change. Monitor authentication errors, throttling, server errors, invalid structured outputs, safety-filter responses, latency, and cost signals.
A rollback can restore your previous application release, but it cannot restore GitHub Models after retirement. The useful rollback target is therefore another functioning provider or deployment, not the retired endpoint. Remove obsolete secrets and dead routing logic only after all consumers are accounted for.
What this means
For a hobby project with no active calls, the correct action may be to archive the integration and document why it no longer runs. For an internal prototype, a Foundry deployment can preserve multi-model experimentation, but someone must own the Azure subscription, permissions, and spend controls. For a production application, the retirement is an infrastructure migration that deserves staged validation and an alternate route.
The main decision is not “Foundry or Copilot?” It is whether the workload needs an inference API, an interactive coding assistant, or both. Copilot can support developers while they rewrite an integration, but it does not serve end-user application requests as a substitute for the retired Models API.
Verification checklist
- No active code, job, notebook, or workflow calls a GitHub Models endpoint.
- Every former model use maps to a deployed and regionally available replacement.
- The application uses the replacement deployment name, endpoint, and supported API version.
- Authentication works in development, staging, and production without credentials in source control.
- Structured outputs, streaming, tools, embeddings, retries, and timeouts are tested where applicable.
- Safety filters and refusal handling match the application’s policy.
- Rate limits, concurrency, and failure behavior are tested under representative load.
- Azure billing ownership, budgets, and alerts are configured before production traffic.
- Logs and metrics exclude sensitive prompt data unless collection is explicitly approved.
- A functioning fallback exists for critical workloads.
Frequently asked questions
When did GitHub Models shut down?
GitHub set July 30, 2026 as the full retirement date. It says the retirement applies to all customers, including those with active usage.
Which GitHub Models features were retired?
GitHub’s notice names the playground, model catalog, inference API, and BYOK endpoints. The related user interface is also removed.
Is GitHub Copilot a replacement for the GitHub Models API?
No. GitHub recommends Copilot for AI-powered workflows directly on GitHub, while it points applications needing model access toward Microsoft Foundry. Copilot is not a drop-in general inference endpoint for an application backend.
Can I keep the same code when moving to Microsoft Foundry?
Microsoft says its guided upgrade can work by replacing the key and endpoint, but Foundry requires explicit model deployments and uses deployment names in requests. Validate parameters, outputs, authentication, filters, and operational behavior before treating the change as compatible.
Is Microsoft Foundry free like GitHub Models experimentation was?
Microsoft’s guide says Foundry usage is billed to an Azure subscription according to the chosen deployment type. Actual availability, quotas, and charges can vary by model, region, and deployment, so verify them in the Foundry portal.