GitHub Copilot's HydraFusion, which can run two models on one request, reaches VS Code and the Copilot app

Since September 4, HydraFusion existed only inside Copilot CLI, behind the /experimental command. On September 30, 2026, GitHub made the same research preview available in VS Code 1.140 or later and in the GitHub Copilot app. You pick it where you would pick a model, and from there Copilot decides whether one model solves the task or two take turns.

Three workflows, chosen per request

GitHub describes HydraFusion as treating workflow selection as an optimization problem. It reads capability signals for reasoning, code generation, debugging, and tool use, then picks the cheapest execution pattern that it expects to meet the quality bar. The pattern is one of three.

• Single: one selected model solves the task directly.
• Cascade: an efficient model drafts a solution, and a quality gate either accepts it or escalates to a stronger model.
• Critique: one model drafts, an independent read-only critic from a different model family reviews it, following the same review pattern as Rubber Duck, and the drafting model revises once.

How it differs from Auto

Auto, the existing default, selects a model for each request and stops there. HydraFusion selects a workflow and coordinates more than one model within a single turn. GitHub's September 4 research post says it chooses from models across multiple providers, without naming them; Claude Opus 5 and GPT-5.6 Sol served as the comparison baselines.

GitHub's own controlled evaluations from that post put HydraFusion 4.9 percentage points above the Opus 5 baseline on TerminalBench 2.1, at 67% lower estimated cost. On DeepSWE it was 1.5 points below at 36% lower cost, and on CheckpointBench 0.1 points below at 65% lower cost. These are the vendor's offline numbers, with no independent replication.

Who gets it and how to turn it on

The preview is available on Copilot Pro, Pro+, Business, and Enterprise. On Business and Enterprise an administrator has to allow preview features first. In VS Code you select HydraFusion from the Copilot Chat model picker; if it is missing, you enable chat.copilot.hydraFusion.enabled. In the Copilot app you update, search Settings for HydraFusion, turn it on, and pick it from the model picker.

GitHub says the expansion to more surfaces was the top request from early feedback. The same release, GitHub adds, shows each workflow step more clearly, sends progress updates more often, and keeps the status readable during long tasks. It remains a research preview and is subject to change.

What it costs

The September 30 post says nothing about billing. The September 4 post does: usage is based on the tokens consumed by the models HydraFusion runs, priced at each model's standard rate. When the workflow is Cascade or Critique, one request pays for two models. This suggests the bill per request is harder to predict than with a model you chose yourself, and GitHub's cost reductions are averages across benchmark runs, not a guarantee per task.

The same post states the current limit: HydraFusion is optimized for first-turn, single-prompt coding tasks, with multi-turn improvements planned next. A long back-and-forth in chat is not what it was tuned for yet.