Short answer: GitHub Models is gone, and GitHub is not restoring it. GitHub’s own guidance points to Azure AI Foundry for model access and GitHub Copilot for IDE workflows. If you were using the inference API in an app or CI job, the replacement is another OpenAI-compatible endpoint — Azure, Plugsky, OpenAI, Anthropic, Google or a self-hosted model.
What actually shut down
GitHub’s retirement notice is explicit. As of 30 July 2026, the following are no longer available to any customer:
- The GitHub Models playground.
- The model catalog.
- The inference API.
- Bring your own key (BYOK).
GitHub also states that GitHub Models was a separate service from GitHub Copilot and is unrelated to Copilot. If your only use was Copilot in the editor, nothing changes; if your code called the GitHub Models inference endpoint, it no longer works.
The four migration paths
| Path | Best for | Effort | Trade-off |
|---|---|---|---|
| Azure AI Foundry | Teams already in Azure; GitHub’s named alternative | Medium | Azure identity, billing and region setup; model availability varies |
| GitHub Copilot | IDE assistance, code review, chat | Low | Not a drop-in inference API for your own app |
| Hosted OpenAI-compatible API (Plugsky, OpenAI, Anthropic, Google, Groq) | Apps, agents and CI that need an HTTP endpoint | Low–medium | New key, base URL, model names and rate limits |
| Self-host an open-weight model | Data cannot leave your network; steady high volume | High | GPU, serving stack, operations and upgrades |
If you used GitHub Models for a prototype
This is the easy case. The code is usually an OpenAI-style client with a GitHub token and a model name. Change three things: the base URL, the API key and the model ID. Then run the same prompts you used before and compare the output.
Do not assume a model with a similar name behaves the same. GitHub Models hosted several families, and the replacement may be a different version, a different quantisation or a different context limit. Treat the first run as an evaluation, not a deploy.
The OpenAI migration checker lists the common incompatibilities: base URL, auth header, streaming format, tool-call schema and JSON mode. Fix those before you touch prompts.
If you used it in CI or production
CI is where retired endpoints cause the most damage, because the failure appears as a red build rather than a customer-facing error. Check four places:
- Secrets. Remove the old GitHub token from Actions secrets and add the new provider key. Do not reuse a personal token for a service.
- Model IDs. Pin the exact model version in config. A floating “latest” alias can change under you.
- Rate limits. The old quota is gone. Run the new provider’s limits through the rate-limit calculator before the first large batch.
- Fallback. Add a second provider or a smaller model for non-critical jobs. A single endpoint is a single point of failure.
If the job is a nightly classification or backfill, move it to a batch queue where the provider offers one. If a user is waiting, keep it synchronous and size the concurrency from p95 latency.
Mapping models without guessing
There is no universal mapping table, because the old catalog and the new provider catalogs do not line up one-to-one. Build a short list instead:
- Small, fast classification/extraction: a small hosted model or an 8B-class open-weight model.
- General chat and RAG: a mid-size model with at least 128K context and tool calling.
- Hard reasoning and code: a frontier or reasoning model; budget for more output tokens.
- Long documents: a model with 256K–1M context, but retrieve before you send the whole document. The context-window calculator shows what is actually usable.
Run 20–50 real prompts per candidate and score them on correctness, format compliance and latency. A public benchmark is a shortlist tool, not a decision.
Plugsky as a GitHub Models replacement
Plugsky exposes an OpenAI-compatible API with 39 models, including the six Plugsky tiers and hosted Qwen, Mistral and DeepSeek options. The base URL is https://plugsky.com/v1; the same SDK call works after you change the key and model.
The commercial difference is the bill. Plugsky’s paid plans are flat-rate rather than per-token, with published RPM, concurrency and daily caps. That is useful for CI and agent workloads where the token count is hard to forecast. The pricing page lists the caps; the cost calculator shows the break-even against per-token providers.
If you need a specific model family, check all models before migrating. If you need a region, VPC, on-prem or air-gapped deployment, start with contact.
Migration checklist
- Inventory every call to the old endpoint: app, scripts, CI, notebooks, agents.
- Choose one replacement path. Do not split the first migration across three providers.
- Move keys into a secret manager; delete the old token.
- Change base URL and model ID in config, not in code.
- Run the evaluation prompts and compare output, latency and cost.
- Test rate limits, retries, streaming and tool calls under load.
- Add a fallback and a status check, then remove the old integration.
Test the replacement in 10 minutes
from openai import OpenAI
client = OpenAI(
base_url="https://plugsky.com/v1",
api_key="YOUR_PLUGSKY_KEY",
)
resp = client.chat.completions.create(
model="plugsky-lite",
messages=[{"role": "user", "content": "Return JSON: {\"ok\": true}."}],
)
print(resp.choices[0].message.content)
Run one plain completion, one JSON-mode call and one tool call. If all three pass, move the rest of the workload. If a tool call fails, check the model’s capability list before changing the schema.
Sources
- GitHub Docs — GitHub Models has been retired (retired 30 July 2026).
- Azure AI Foundry — GitHub’s named alternative for model access.
- GitHub Copilot documentation — separate from GitHub Models.
FAQ
Is GitHub Models coming back?
No. GitHub states that GitHub Models was fully retired on 30 July 2026 and that the playground, model catalog, inference API and bring-your-own-key feature are no longer available. GitHub points users to Azure AI Foundry for model access and GitHub Copilot for IDE workflows.
What is the closest replacement for the GitHub Models API?
The closest replacement is another OpenAI-compatible inference API. Azure AI Foundry is the option GitHub names. Plugsky, OpenAI, Anthropic, Google and Groq are also options; the right one depends on model choice, region, rate limits and whether you need a flat bill.
Does this affect GitHub Copilot?
No. GitHub says GitHub Models was a separate service from GitHub Copilot and is unrelated to Copilot. Copilot continues under its own plan and documentation.
How long does migration take?
For a small prototype, an afternoon: change the base URL, key and model name, then run your evaluation prompts. For CI or production, allow a sprint for secrets, model mapping, rate-limit testing, fallback and cost review.
Need an OpenAI-compatible endpoint after GitHub Models?
Plugsky has 39 models, a free tier and flat-rate paid plans. Change the base URL and key, then run your evaluation.
Start Free → See plans