GitHub HydraFusion in VS Code: Enable Copilot’s Multi-Model Coding Preview

The launch arrived on schedule. When GitHub first announced the project, its community thread said HydraFusion was "Today, only in GitHub Copilot CLI. It shipped in the latest release as an experimental research preview. The GitHub Copilot app and VS Code are both targeting September as a fast follow." September 30 is the last day of September, so GitHub made its own deadline with very little room to spare.

What HydraFusion actually is (and isn't)​

HydraFusion appears in Copilot's model picker, but it is not a model. It is a runtime orchestrator. For each prompt, it picks an execution pattern and the models that run it, then returns one answer. Next AI Press put it this way: HydraFusion isn't a new AI model. It's an orchestration layer that sits on top of GitHub Copilot's existing model lineup and decides, per request, how many models to involve and in what order.

GitHub describes workflow selection as an optimization problem. HydraFusion weighs capability signals for reasoning, code generation, debugging and tool use, then picks the most efficient pattern it expects to meet its quality bar. There are currently three patterns:

• Single: One model handles the task directly.
• Cascade: A cheaper, faster model writes a draft. A quality gate then accepts it or escalates the task to a stronger model.
• Critique: One model drafts. A read-only reviewer from a different model family critiques the draft, and the first model revises once. GitHub compares this to the "rubber duck" review agent in Copilot CLI.

According to GitHub's documentation, the choice is made per prompt. Your first request might run as Single and a follow-up as Critique. HydraFusion adds extra model passes only when it judges a task needs them. Cascade and Critique take longer because they include those passes.

The short version: HydraFusion is a workflow picker. It sits in the model dropdown, and the extra review or escalation steps happen without you managing them.

HydraFusion vs. Auto: which one to use​

Copilot already has Auto model selection, so the obvious question is how the two differ. GitHub's community FAQ answered it: Auto selects one model per request. HydraFusion runs multiple models in a single turn, has them review and critique one another, and fuses the result. Auto picks the optimal model; HydraFusion picks the optimal workflow, possibly using multiple models.

The same thread describes the lineage: Auto V1 (Jan 2026, capacity/SKU-aware per-request selection) → Auto V2 aka. HyDRA (May 2026, intent-scored routing) → HydraFusion (Aug 2026, per-turn orchestration and cache-aware workflows).

GitHub's own guidance on when to use each:

Setting it up in VS Code​

You need VS Code 1.140 or later, or VS Code Insiders. GitHub's documented steps:

1. Open the Settings editor with Ctrl+, on Windows or Linux (Command+, on Mac).
2. Search for chat.copilot.hydraFusion.enabled and tick the checkbox.
3. Open Copilot Chat from the chat icon in the VS Code title bar.
4. At the bottom of the chat view, open the model dropdown and choose HydraFusion. It appears just below Auto.
5. Enter a prompt. The chat view shows each step of the pattern as it runs.
6. To see which models were used, hover over the footer of the completed response.

The September 30 changelog says to try the model picker first and turn on the setting only if HydraFusion isn't listed. Enabling it first, as the documentation does, reaches the same result.

The setting has been around for a few weeks. Nerd Level Tech reported that on September 11, VS Code merged an off-by-default, experimental chat.copilot.hydraFusion.enabled setting for its Copilot agent host into its main development branch, though a cherry-pick into the 1.138 release branch was closed unmerged. That explains why the documentation sets 1.140 as the minimum.

Setting it up in the GitHub Copilot app​

1. Update the Copilot app to the latest version.
2. Open Settings, then Experimental. The changelog also says you can search Settings for "HydraFusion".
3. Turn HydraFusion on.
4. Choose HydraFusion in the model picker.
5. Hover over a response to see which models it used.

If the setting isn't there after updating, GitHub's troubleshooting guide says to switch to the app's prerelease channel in its settings.

For comparison: Copilot CLI​

The CLI setup hasn't changed, but it's worth knowing because the CLI still has the most detailed progress view and diagnostics. Run copilot update, start the CLI with copilot --experimental (or enter /experimental on in a session and restart), then use /model to choose HydraFusion (Research Preview).

For scripts, GitHub documents a single-prompt option: copilot --experimental --model hydrafusion -p "YOUR-PROMPT". The CLI also offers /usage for per-model AI credit use, and /collect-debug-logs for a record of which pattern and models each step used. GitHub suggests including those logs in bug reports.

Who can use it​

The changelog lists Copilot Pro, Pro+, Business, and Enterprise. For Business and Enterprise, an administrator must turn on preview features at the organization or enterprise level.

That list is narrower than at launch. In early September, GitHub's thread said HydraFusion is available to users on all GitHub Copilot plans through /experimental in GitHub Copilot CLI. The September 30 changelog doesn't mention Copilot Free. Free-tier users shouldn't count on seeing HydraFusion in VS Code or the app until GitHub says otherwise.

Model policies also matter. HydraFusion uses only models that your plan includes and that your organization or enterprise allows. You can't choose its models yourself. If policy blocks every model it would use, HydraFusion doesn't appear in the picker at all.

If HydraFusion doesn't appear​

• VS Code: Check that you're on 1.140 or later and that chat.copilot.hydraFusion.enabled is on.
• Copilot app: Update, check Settings > Experimental, and try the prerelease channel if the toggle is missing.
• Managed accounts: Ask your admin whether preview features are enabled for you and whether your model policies allow any of the models HydraFusion uses.
• Asked to compact the conversation: HydraFusion's displayed context window is based on the smallest limit among its models. If your current chat is larger than that, compact it before switching.

What it costs, and the catches​

HydraFusion has no separate fee. You pay for each model it uses at that model's standard rate, and the Auto discount doesn't apply. A Critique run bills a drafting model and a reviewer, so one task can use more AI credits than a single-model request. GitHub says it keeps the main conversation on one model where possible to benefit from cached tokens, and helper models such as reviewers get only the context they need.

Two limitations matter most in practice:

• Discarded drafts don't undo their edits. If HydraFusion throws away an intermediate draft, file changes that draft already made are not rolled back automatically. GitHub says to review changes before committing. Treat HydraFusion output like any other agent change set.
• It isn't for production work. GitHub offers no SLA during the preview and says HydraFusion is not intended for production workloads. The patterns and models can change at any time. It also doesn't start subagents.

The preview may also suit some kinds of work better than others. At the CLI launch, Next AI Press reported that GitHub says it currently works best on first-turn, single-prompt coding tasks — the kind you'd hand to Copilot in autopilot mode and let it complete in one go. GitHub's current documentation doesn't repeat that restriction. Even so, if a long, back-and-forth session gives uneven results, this is a likely reason.

What the benchmarks show​

GitHub's research post, published September 4, reports controlled offline evaluations on three agentic coding benchmarks against Claude Opus 5. Pondero summarized them: Task quality on TerminalBench came in 4.9 percentage points above Opus 5. On DeepSWE, cost reduction reached 36% with quality 1.5 percentage points below Opus 5. TerminalBench 2.1 also showed 67% lower estimated cost. On GitHub's internal CheckpointBench, cost was 65% lower and quality was 0.1 points below Opus 5.

So HydraFusion clearly beat Opus 5 on one benchmark and came close on the other two, at much lower estimated cost. GitHub notes that these results apply only to the tested benchmark versions, configurations, model pool and pricing assumptions, with every model at medium reasoning. GitHub runs the benchmarks, the product and the billing. As Pondero noted, GitHub has not announced a timeline for HydraFusion exiting research preview or for production cost data. Your own repository and your AI credit usage are the better test.

Is it worth trying?​

For developers on Windows who spend the day in VS Code, the barrier to trying HydraFusion is now one setting and a dropdown choice. A reasonable approach:

• Keep Auto as your default.
• Use HydraFusion for bounded, harder tasks where a second model's review is worth some extra time and credits, such as a stubborn bug or a coordinated change across several files.
• Watch your AI credit usage for the first week, especially on tasks that run as Cascade or Critique.
• Review every diff before committing, since a discarded draft can leave edits behind.
• Admins: Decide on preview-feature and model policies before developers start asking why HydraFusion isn't in their picker.

HydraFusion now runs in all three places GitHub promised, and the progress display is clearer than before. Whether it saves money on real workloads, as it did on GitHub's benchmarks, is still unproven. The preview is the place to find out.