GitHub has expanded HydraFusion beyond Copilot CLI to Visual Studio Code and the GitHub Copilot app. The feature is available as a research preview for eligible Copilot plans and can be selected from the familiar model picker, but it is not a conventional single-model choice. HydraFusion coordinates multiple models and selects an execution workflow for a coding request.
That distinction matters for developers evaluating coding quality, response time, model usage, and control. GitHub describes HydraFusion as an experimental system intended to balance those factors, while also warning that its models, workflows, availability, names, and behavior may change.
What changed
According to GitHub’s September 30, 2026 announcement, HydraFusion is now available in Visual Studio Code and the GitHub Copilot app, in addition to Copilot CLI. Users select it from the Copilot model picker, although the selection represents an orchestration system rather than one underlying model.
HydraFusion can select models across multiple providers and uses capability signals related to reasoning, code generation, debugging, and tool use. GitHub has not published the complete model pool, routing thresholds, or the full policy used to make those choices. Users also are not documented as having a control to manually choose or override the workflow selected for an individual request.
The system currently documents three workflow patterns:
- Single: One selected model works on the task directly.
- Cascade: An efficient model produces a draft, then a quality gate either accepts it or escalates the request to a stronger model.
- Critique: A model drafts the result, an independent model from a different family reviews it in a read-only context, and the drafting model performs one revision.
The Critique pattern resembles GitHub’s Rubber Duck agent review process. The reviewer can examine plans, code, or tests without changing files or altering the environment. GitHub says HydraFusion review steps run in isolated, tool-less contexts, while solver steps use the shared workspace and a permission-aware agent loop.
How to enable HydraFusion
Visual Studio Code
VS Code users need version 1.140 or later, or VS Code Insiders. Open Copilot Chat and select HydraFusion from the model picker.
If it does not appear, GitHub instructs users to enable the following setting:
chat.copilot.hydraFusion.enabled
The announcement does not provide a broader compatibility matrix beyond the stated VS Code requirements.
GitHub Copilot app
For the GitHub Copilot app, update to the latest available version, open Settings, search for “HydraFusion,” and enable it. HydraFusion can then be selected from the app’s model picker.
GitHub has not documented the exact required app version or a complete list of supported app platforms in the supplied release information. That means teams should verify the option in their installed app rather than assume that every Copilot app installation has identical support.
Plans and organization controls
The announcement lists Copilot Pro, Pro+, Business, and Enterprise users as eligible. For Business and Enterprise organizations, administrators must allow preview features where required. GitHub’s policy documentation notes that preview features can contain flaws and may be changed or discontinued.
What the orchestration means for coding work
HydraFusion may use different execution patterns for different tasks instead of sending every request to one fixed model. A straightforward request may use the Single workflow. A task that benefits from a quality check may enter Cascade or Critique, potentially adding additional model calls and execution steps.
GitHub says the runtime is designed to account for each workflow leg, enforce bounded execution with timeout and cancellation behavior, isolate review contexts, validate routing and model availability before execution, and apply patches safely. It also says that no patch is applied when a workflow is cancelled or fails validation. Runtime records include each leg’s role, outcome, cost, latency, and diagnostics.
These are documented design principles, not a guarantee that generated code is correct or secure. Developers still need to inspect changes, run tests, and maintain normal code review and security practices. The Critique workflow should not be treated as a replacement for human review.
GitHub has also improved visibility into workflow steps and added more frequent progress updates, including clearer status during longer tasks. This may make multi-step execution easier to monitor, but the supplied information does not establish live latency or reliability for the VS Code and Copilot app rollout.
What GitHub’s benchmark results do—and do not—show
GitHub’s HydraFusion research post reports offline evaluations using TerminalBench 2.1, DeepSWE, and CheckpointBench. It cites a TerminalBench 2.1 result that was 4.9 percentage points higher in verified task quality and had an estimated cost 67% lower than Claude Opus 5.
Those figures are narrowly scoped. GitHub ties them to the evaluated benchmark revisions, workflow configurations, model pool, pricing assumptions, and reasoning settings. They do not demonstrate that the same quality, cost, or latency will occur in live Visual Studio Code sessions or in the GitHub Copilot app.
GitHub recommends beginning with substantial, well-scoped, first-turn single-prompt coding tasks. Performance in longer, multi-turn iterative sessions remains a future focus for the preview.
Important limitations for teams
HydraFusion is still an active research preview. The complete model pool, routing algorithm, quality-gate thresholds, fallback behavior, and user workflow override controls are undocumented. The sources also do not establish whether Single, Cascade, and Critique are exposed identically across VS Code, the Copilot app, and Copilot CLI.
Billing behavior is another unresolved detail. GitHub describes accounting for workflow legs and diagnostics, but the supplied documentation does not establish HydraFusion-specific credit consumption, pricing, or whether all internal legs are billed identically across plans.
Privacy and data routing also require caution. General Copilot documentation describes provider-specific hosting, retention, and session behavior, but it does not identify which provider handles each HydraFusion request or provide a complete HydraFusion-specific retention and training policy. Organizations should avoid treating general Copilot policy information as a complete description of HydraFusion’s data path.
What you should do
- Update VS Code to 1.140 or later, or use VS Code Insiders.
- Select HydraFusion in Copilot Chat. If it is missing, enable
chat.copilot.hydraFusion.enabled. - Update the GitHub Copilot app to its latest version, enable HydraFusion in Settings, and select it from the model picker.
- If you use Copilot Business or Enterprise, confirm that the organization allows the relevant preview features.
- Start with focused, first-turn coding tasks rather than assuming the preview is optimized for long iterative sessions.
- Review generated changes, run appropriate tests, and do not use the orchestration safeguards as a substitute for engineering review.
For now, HydraFusion is best viewed as an experimental multi-model coding workflow. It may offer a different balance of quality and resource use, but its live behavior, billing, routing, and data-handling details remain subject to change or further documentation.
