GitHub Project HydraFusion: How Copilot’s Multi-Model Orchestration Works

01

What GitHub Project HydraFusion is

GitHub introduced Project HydraFusion on September 4, 2026 as a research preview inside GitHub Copilot CLI. Rather than binding a request to one model, HydraFusion can choose a compound execution plan using models from multiple providers.

GitHub frames the system as runtime orchestration: the developer selects HydraFusion, while the service decides which workflow is most likely to meet a quality target while balancing cost and latency. It is not a new foundation model.

02

Single, Cascade and Critique workflows

HydraFusion currently chooses among three patterns. Single uses one selected model. Cascade starts with a more efficient model and escalates only when a quality gate rejects the first result. Critique has one model draft, another model family review in an isolated read-only context, and the original drafting model revise once.

The preview automatically selects the workflow for each request. Developers are not choosing Single, Cascade or Critique manually for every prompt in the launch implementation.

03

Availability and launch setup

As of September 5, 2026, GitHub says HydraFusion is available to users on all GitHub Copilot plans through the experimental feature path in Copilot CLI.

1. Run <code>/update</code> in Copilot CLI.
2. Run <code>/experimental on</code>.
3. Run <code>/model</code> and select <code>HydraFusion (Research Preview)</code>.

04

How HydraFusion billing works

GitHub says usage is based on the tokens consumed by the underlying models HydraFusion invokes, with those tokens priced at each model's standard rate. There is no published separate HydraFusion orchestration surcharge in the launch announcement.

A Critique or Cascade workflow can contain multiple inference legs, so a single request can consume usage across drafting, review, revision, escalation, retries or fallbacks. The system is designed to reduce expensive calls where possible, not to guarantee that every request is cheaper than one direct model call.

05

GitHub's benchmark claims and what they mean

GitHub reports controlled offline evaluations on TerminalBench 2.1, DeepSWE and CheckpointBench. Relative to Claude Opus 5, GitHub reports 4.9 percentage points higher verified quality at 67% lower estimated cost on TerminalBench 2.1, 1.5 points lower quality at 36% lower cost on DeepSWE, and 0.1 points lower quality at 65% lower cost on CheckpointBench.

06

Repository-state and execution controls

GitHub describes complete accounting, bounded execution, isolated review, fail-safe patch application and validated routing as core operating principles. Critic steps are tool-less, while solver steps remain in the normal permission-aware agent loop.

HydraFusion also withholds intermediate drafts and returns one coherent result because some intermediate work may later be revised, escalated or discarded. That improves final-state clarity but can make long compound workflows feel less transparent while they run.

07

Who should try it and current limitations

The preview is most relevant to developers already using Copilot CLI who want to test cost-aware routing, model-independent critique and compound coding workflows. GitHub currently recommends substantial, well-scoped first-turn coding tasks as a good starting point.

HydraFusion remains a research preview. Models, commands, routing behavior, latency, availability and billing details can change, and teams with strict production guarantees should validate it on their own repositories before making it a default path.