HydraFusion no longer requires a terminal: preview available in VS Code and the Copilot app

Previously, the preview was available only in GitHub Copilot CLI. Now you can select HydraFusion from the model menu in VS Code and the GitHub Copilot app. It is a mode for using models within Copilot, rather than a separate service.

Auto selects one model per request. HydraFusion chooses one of three approaches:

- Single: one model handles the task. - Cascade: a stronger model continues the work if needed. - Critique: another model independently reviews the answer, followed by one revision.

How to enable it. In VS Code 1.140+ or Insiders, select HydraFusion in Copilot Chat. If it is missing from the menu, enable `chat.copilot.hydraFusion.enabled`.

In the Copilot app, update to the latest version, open Settings → Experimental and enable HydraFusion. Then select it from the model menu. If the toggle is missing, try the prerelease channel in settings.

In the CLI, run `copilot update`, launch `copilot --experimental` and select HydraFusion (Research Preview) via `/model`. For scripts, use `copilot --experimental --model hydrafusion -p "YOUR-PROMPT"`.

The preview is available on these plans:

- Copilot Pro. - Copilot Pro+. - Copilot Business. - Copilot Enterprise.

For Business and Enterprise, an administrator must allow preview features.

How usage is billed. There is no separate charge for HydraFusion, but all models used are billed at standard rates. The Auto discount does not apply. One task may use more AI credits than running a single model.

In GitHub's offline evaluations from Sep 4, HydraFusion cost less than Claude Opus 5. Quality improved on one benchmark and declined slightly on two:

| Benchmark | Quality change relative to Claude Opus 5 | Change in estimated cost | | --- | --- | --- | | TerminalBench 2.1 | +4.9 percentage points | −67% | | DeepSWE | −1.5 percentage points | −36% | | CheckpointBench | −0.1 percentage points | −65% |

In kkamegawa's experiment on Sep 6, HydraFusion independently published a NuGet package and moved a draft to release, meeting the issue's acceptance criteria. Completing the entire issue cost about 1,000 credits.

GitHub has added more frequent progress updates. You can see which models were used by hovering over the bottom line of the response in VS Code or over the response in the Copilot app.

In the CLI, the `/usage` command shows usage for each model used.