🐉 Using HydraFusion Across Copilot Surfaces

Picking a model used to be the last decision you made before pressing Enter. Then Auto model selection took that decision away for everyday work. HydraFusion goes one step further: instead of picking one model, it picks a whole workflow, deciding whether a task needs one model, an escalation to a stronger one, or a draft that other models critique before you see it.

HydraFusion is documented under GitHub’s early-access docs and is currently a research preview. That means no SLA, behaviour that can change, and a clear statement from GitHub that it is not intended for production workloads. It is still very usable for real coding tasks today, as long as you understand how it spends AI credits and where it is (and is not) available.

This post covers what HydraFusion is, how to enable it on each Copilot surface, and the limitations worth knowing before you hand it a large task.

Last verified

HydraFusion changes quickly, so this is a living guide. An automated check compares it with the official documentation and the release notes of each tool, and every update is reviewed by a human before it is published. The table shows which versions the steps were last checked against.

Does a step no longer match what you see in your version? Report an outdated step and include your tool version.

What HydraFusion is

HydraFusion is a runtime model orchestrator in GitHub Copilot. You select it in the model picker, but it is not a model. For each request, it chooses an execution pattern, the reasoning effort for each step, and the models that run it, then returns a single response.

It currently chooses between three execution patterns:

A few behaviours are worth calling out:

• Per-prompt decisions: the pattern is chosen for every prompt, so one session can use Single for a first prompt and Critique for a follow-up.
• Lightweight selection: choosing the pattern adds little time and does not generate any part of the response.
• Per-step reasoning: drafts run with medium reasoning and reviews with high reasoning. HydraFusion Max unlocks extra high reasoning.
• No fixed model list: the model mix changes as new models arrive and GitHub re-evaluates performance. You cannot choose which models it uses.
• Final answer only: you can follow the progress of each step, but intermediate drafts are hidden because they may be revised or discarded.
• No context window of its own: each step runs within the limits of the model running it. The context window shown for HydraFusion is a conservative value based on the smallest limit among its models.

The Critique pattern will feel familiar if you have used the rubber duck agent in Copilot CLI, which deliberately runs a critic on a different model from your session model to avoid shared blind spots.

What the benchmarks say

GitHub’s Project HydraFusion blog post reports controlled offline evaluations of the best tuned configuration against Claude Opus 5:

GitHub is explicit that these results are specific to the evaluated benchmark revisions, model pool, and pricing assumptions. Treat them as a signal of intent rather than a guarantee for your repository.

The same post describes the operating principles behind the runtime, including isolated, tool-less review steps and fail-safe application: no patch is applied when a workflow is cancelled or fails validation.

HydraFusion vs Auto

Both options make the choice for you, but they choose different things. Auto picks the best model per request. HydraFusion picks the best execution pattern and reasoning effort per task, possibly using several models.

The GitHub blog recommends starting with first-turn, single-prompt coding tasks that you can hand to Copilot in autopilot mode. Strong multi-turn performance for longer, iterative sessions is what GitHub says it is focusing on next.

Prerequisites and access

Before you go looking for HydraFusion in a model picker, check the following:

• A Copilot plan: the announcement states HydraFusion is available to users on all GitHub Copilot plans.
• A supported surface: Copilot CLI, Visual Studio Code, or the GitHub Copilot app.
• Preview features allowed: if you get Copilot through an organisation or enterprise, your administrator must allow preview features for you.
• Model policies: HydraFusion only uses models available in your plan and allowed by your organisation or enterprise model policies. If none of its models are available, HydraFusion does not appear in the picker.

Billing

There is no separate HydraFusion charge. You are billed for each model it uses at that model’s standard rate, and the Auto model selection discount does not apply. Because Cascade and Critique run more than one pass, a single task can consume more AI credits than the same task on one model.

To keep costs reasonable, HydraFusion keeps the main conversation on the same model whenever possible so cached tokens continue to apply, and assisting models such as reviewers receive only the context they need. Per-model rates are in Models and pricing for GitHub Copilot, where 1 AI credit equals $0.01 USD.

If you are already tracking Copilot spend, the guidance in GitHub Copilot FinOps applies directly here.

Using HydraFusion in Copilot CLI

Copilot CLI was the first surface to ship HydraFusion and has the richest progress display.

1. Update Copilot CLI to the latest version:

   copilot update
   

2. Start Copilot CLI with experimental features turned on:

   copilot --experimental
   

   If you are already in a session, enter /experimental on and then restart Copilot CLI.

3. Enter /model and select HydraFusion (Research Preview) or HydraFusion Max (Research Preview).

4. Enter a prompt.

While HydraFusion works, Copilot CLI shows:

• The execution pattern it chose: Single, Cascade, or Critique
• Each planned pass and whether it is done, running, or pending
• What the current pass is doing, such as running tools or receiving output
• How long the task has been running

Press Esc to interrupt. When the task completes, the conversation keeps a summary of the execution pattern and its steps, including any warnings.

Non-interactive use

For a single prompt, for example in a script, pass the model with --model:

copilot --experimental --model hydrafusion -p "Fix the flaky retry logic in src/http/client.ts and add a regression test"

Checking usage and collecting logs

Two slash commands are useful during the preview:

• /usage shows the AI credits each model consumed in the current session.
• /collect-debug-logs creates an archive of debug logs with the execution pattern and models used per step. Attach this when filing a bug report.

Using HydraFusion in Visual Studio Code

HydraFusion shipped in VS Code 1.140 as a research preview. You need VS Code 1.140 or later, or VS Code Insiders.

1. Open the Settings editor with Ctrl+, on Windows and Linux or Cmd+, on macOS.
2. Search for chat.copilot.hydraFusion.enabled and select the checkbox.
3. Open Copilot Chat from the chat icon in the title bar.
4. At the bottom of the chat view, open the current model dropdown and select HydraFusion or HydraFusion Max.
5. Enter a prompt. The chat view shows each step of the execution pattern as HydraFusion works.

If you manage settings as code, the equivalent settings.json entry is:

{
  "chat.copilot.hydraFusion.enabled": true
}

To see which models HydraFusion used, hover over the footer of a completed response.

Using HydraFusion in the GitHub Copilot app

The GitHub Copilot app exposes HydraFusion through its experimental settings:

1. Update the Copilot app to the latest version.
2. Open Settings and select Experimental.
3. Turn on HydraFusion.
4. In the model picker, select HydraFusion or HydraFusion Max.

To see which models HydraFusion used, hover over the response. If the HydraFusion setting is missing, switch to the prerelease channel in the app’s settings.

What about Visual Studio?

At the time of writing, GitHub’s HydraFusion documentation lists availability in Copilot CLI, Visual Studio Code, and the GitHub Copilot app only. Visual Studio is not listed, and neither the Visual Studio Copilot documentation nor the Visual Studio 2026 release notes mention HydraFusion.

If you work primarily in Visual Studio, the practical options today are:

• Use Auto model selection in Visual Studio for everyday work.
• Run larger, well-scoped HydraFusion tasks from Copilot CLI in a terminal against the same repository, then review the changes in Visual Studio.

Keep an eye on the HydraFusion docs page, since the availability list is the authoritative source if support expands.

Comparing the surfaces

Limitations and tips

The documented limitations are short, but each one matters in practice:

• No subagents: HydraFusion works separately from subagents and does not start them.
• Discarded drafts can leave edits behind: if HydraFusion discards a draft, file edits that draft already made in your workspace are not undone automatically. Review your working tree before committing.
• Moving target: execution patterns and models can change during the research preview.

A few practical tips based on the documentation:

• Use a clean branch or worktree for HydraFusion runs so any leftover edits from discarded drafts are easy to spot with git status and git diff.
• Reserve it for substantial tasks. For quick or routine prompts, switch back to Auto. Cascade and Critique take longer and cost more.
• Compact before switching in long sessions. If your conversation is larger than HydraFusion’s conservative context window, you are asked to compact first. See Managing context in GitHub Copilot CLI.
• Watch credits early. Check /usage in the CLI or monitor your AI Credits usage on GitHub after the first few runs to calibrate expectations.

Troubleshooting

HydraFusion does not appear in the model picker

• Copilot CLI: confirm experimental features are on and restart after /experimental on. If it is still missing, enter /update prerelease. Prereleases may be less stable.
• VS Code: confirm you are on 1.140 or later and chat.copilot.hydraFusion.enabled is on.
• Copilot app: confirm you are on the latest version and HydraFusion is on in Settings > Experimental. Switch to the prerelease channel if the setting is missing.
• Organisation or enterprise: ask your administrator whether preview features are enabled, and whether your model policies allow any of the models HydraFusion uses.

A task is slow or uses more AI credits than expected

Cascade and Critique run multiple passes by design. Use the CLI progress display to see which pattern is running and select Auto for routine work.

You are asked to compact before switching

Your conversation exceeds HydraFusion’s conservative context window. Compact the conversation, then switch.

Recap and next steps

HydraFusion moves Copilot from “choose the best model” to “choose the best way to solve this task”. In practice:

• Enable it in Copilot CLI with --experimental, in VS Code 1.140+ with chat.copilot.hydraFusion.enabled, or in the Copilot app under Settings > Experimental.
• Visual Studio is not currently listed as a supported surface.
• Use it for substantial, well-scoped tasks, and keep Auto for everyday prompts.
• Expect higher credit consumption on multi-pass patterns and review your working tree before committing.

Because this is a research preview, GitHub is actively asking for feedback. Share your results, including session IDs or debug logs, in the HydraFusion feedback discussion or through /feedback in Copilot CLI.

Change log

• 2026-10-16: First published. Added the “Last verified” table and automated release-note checks.

References

• Using HydraFusion (GitHub Docs, early access)
• Project HydraFusion: frontier quality via multi-model orchestration (GitHub Blog)
• HydraFusion research preview announcement and FAQ (GitHub Community)
• VS Code 1.140 release notes: HydraFusion model orchestration
• About Copilot auto model selection
• Getting a second opinion with the rubber duck agent
• Models and pricing for GitHub Copilot
• Monitoring your GitHub AI Credits usage
• Managing context in GitHub Copilot CLI
• Related: Maximising Copilot token efficiency