GitHub announced on September 30, 2026 that HydraFusion is coming to Visual Studio Code and the Copilot app, a month after its launch as a research preview in Copilot CLI. HydraFusion is not an additional model but an orchestration system: for each task, it decides whether to use a single model, escalate to a more capable one, or run a cross-review between two models from different families. The move follows a string of Copilot projects shipped in recent weeks, including the opening of Copilot Code to non-technical roles.
Three ways to run a task
GitHub frames HydraFusion as a workflow optimization problem rather than a plain model picker. Three execution patterns are available, chosen automatically based on the nature of the request.
| Mode | How it works | Typical use case |
|---|---|---|
| Single | One model handles the request end to end | Simple, well-scoped tasks |
| Cascade | A cost-efficient model drafts a response; a quality gate escalates to a stronger model if needed | High volume, constrained token budget |
| Critique | One model drafts, a second model from a different family reviews independently, the first model revises once | Sensitive tasks where a mistake is costly |
From the CLI to the editor
Until now, activation only went through the Copilot CLI command line, via an experimental setting and a dedicated model picker.
/update
/experimental on
/model
> HydraFusion (Research Preview)The extension to VS Code and the Copilot app brings the same mechanism to a graphical interface, inside the regular model picker, without touching the terminal. For teams already running their agents from the command line, nothing changes; for everyone else, it is the first chance to try the feature without leaving the editor. GitHub also says it improved the process’s readability: real-time progress and clearer status indicators during long-running tasks, a weak point frequently reported since the CLI launch.
Gains claimed by GitHub
On its own TerminalBench 2.1 evaluation, GitHub reports an estimated 67% cost reduction compared to running Claude Opus 5 alone, alongside a 4.9 percentage point increase in correctly completed tasks. These figures come from an internal GitHub protocol measured against its own benchmark; they have not been independently verified by this newsroom.
HydraFusion does not replace Copilot’s “Auto” model selector: it explores how a single turn can both pick a workflow and coordinate several models at once.
Access still requires a Copilot Pro, Pro+, Business, or Enterprise subscription. For enterprise accounts, an administrator must enable preview features at the organization level before a developer can use it, whether through the CLI, VS Code, or the Copilot app.
Key takeaways
HydraFusion moves from a CLI-only experiment to broader availability in VS Code and the Copilot app, without changing its core principle: automatically routing a task to a single model, a cost-efficient cascade, or a cross-review. It remains a research preview, gated by organization policy, and the claimed gains rest on a GitHub-run benchmark. It fits a broader push to standardize how agents and tools talk to each other, an area where the MCP specification sets the protocol-level groundwork.
The Cascade mode seems the most likely to see production adoption, because it answers a concrete budget problem rather than an abstract quality promise. Critique, on the other hand, looks most useful on security- or billing-related code, where a second pair of eyes genuinely changes the outcome. I would stay cautious about the cost figures GitHub is citing until an independent party reproduces the measurement: a vendor grading its own feature on its own benchmark rarely comes out behind. — Simon Janvier
Further reading: official announcement on the GitHub changelog.
