GitHub Copilot adds semantic issue search, smarter auto model routing, Eclipse transparency, and web model cleanup

GitHub’s May 20-21, 2026 Copilot updates show GitHub Copilot becoming more surface-specific and governance-aware.

The headline changes: semantic issue search in Copilot Chat on web, auto model selection in VS Code, GitHub-owned download URLs for Copilot usage metrics reports, open-sourcing Copilot for Eclipse, and a cleanup of the model list for Copilot Chat on GitHub.com.

The most important nuance: GitHub removed all Gemini models from Copilot Chat on the web on May 20, even though Gemini 3.5 Flash had just become generally available in GitHub Copilot on May 19. Model access is now clearly surface-dependent.

What changed

Semantic issue search lets users ask Copilot Chat on web to find, group, and analyze issues using natural language. GitHub says it uses a semantic issues index, so a user does not need to remember the exact title or keyword.

Auto model selection in VS Code now routes tasks based on real-time model availability, reliability signals, and task characteristics such as reasoning, code complexity, bug diagnosis, and tool orchestration. GitHub says users can see which model was used, switch away from Auto, and rely on admin policies being respected.

For enterprise operations, Copilot usage metrics report links now use GitHub-owned domains, including copilot-reports.github.com for GitHub.com customers. That is a small change with real security-team value because firewall and proxy allowlists become less brittle.

GitHub also open-sourced Copilot for Eclipse under the MIT license. Developers can inspect implementation details for completions, chat, agent mode, skills and prompt files, bring-your-own-key handling, custom agents, subagents, plan agent behavior, and MCP integration.

Finally, GitHub limited Copilot Chat on GitHub.com to a smaller web model list, removing all Gemini models and several other models from that specific surface to improve response consistency.

Why this matters

Copilot is no longer just autocomplete plus chat. It is becoming a GitHub-native work layer for code, issues, models, reports, IDEs, and cloud agent workflows.

That complexity creates value and risk. Semantic issue search can speed planning and triage. Auto model selection can reduce waste on simple tasks and improve reliability. Usage metrics report stability helps enterprise admins. Eclipse source transparency builds trust for a mature IDE ecosystem.

But the model cleanup is the reminder buyers need: a model being “in Copilot” does not mean it is in every Copilot surface. Web chat, VS Code, JetBrains, Eclipse, CLI, mobile, cloud agent, Business, Enterprise, and Pro+ may not expose the same model behavior.

Buyer take

If your team uses GitHub Issues heavily, semantic issue search is an immediate productivity test. Try it on issue triage, duplicate detection, release planning, platform-specific bug clusters, and support backlog analysis.

If your team burns premium requests, test Auto in VS Code on routine coding and debugging tasks. The best use is not blindly trusting routing. It is comparing output quality and request usage against your manual model picks.

If you are an enterprise admin, update allowlists for the new Copilot report domains and document which Copilot surfaces your model policy actually controls.

What to watch next

Watch whether GitHub publishes clearer per-surface model availability docs and whether model removals become more common as Copilot optimizes for reliability and cost.

The practical rule for buyers: treat Copilot as a platform with multiple AI surfaces, not one product with one model picker.