What Is GitHub's HydraFusion and How Does It Pick Models for Your Code?

How HydraFusion Chooses a Workflow

You open Copilot CLI, run /experimental on, then /model and pick HydraFusion (Research Preview).

From that moment the system treats every request as an optimization problem. It reads capability signals for reasoning, code generation, debugging, and tool use. Then it selects one of three execution patterns.

Single sends the task to one model. Cascade lets an efficient model draft first; a quality gate decides whether to accept or escalate to a stronger model. Critique adds an independent read-only reviewer from a different model family — the same review pattern used in Rubber Duck — then lets the drafter revise once. Each pattern targets a different quality-to-cost trade-off.

The runtime records role, outcome, cost, latency, and diagnostics for every leg. You see one coherent response and one permission-aware change set. Intermediate drafts stay hidden until the workflow finishes. GitHub says showing them live could make unfinished work appear final.

What the Benchmarks Actually Show

Fixed HydraFusion policies ran against three agentic coding benchmarks: TerminalBench 2.1, DeepSWE, and CheckpointBench. Each policy used the same task inputs, tools, execution limits, pricing assumptions, and grading conditions. Cost accounting included every invoked leg: drafting, critique, revision, escalation, retry, and fallback.

On TerminalBench 2.1 HydraFusion improved verified task quality by 4.9 percentage points at 67% lower estimated cost compared with Claude Opus 5. DeepSWE came within 1.5 points of Opus 5 while reducing cost by 36%. CheckpointBench landed within 0.1 points at 65% lower cost. Early internal testing echoed that result.

These controlled offline results are specific to the evaluated benchmark revisions, workflow configurations, model pool, and pricing assumptions. All models ran at the same medium reasoning level. The research preview exists to validate how these results translate to real developer workloads.

What Happens When You Turn It On

Usage is based on the tokens consumed by the models HydraFusion uses, priced at each model's standard rate. No extra markup.

The preview works best with substantial, well-scoped coding tasks you can hand to Copilot in autopilot mode in a single prompt. Multi-turn performance comes next.

Five operating principles govern execution. Complete accounting aggregates cost across every workflow leg. Bounded execution gives each leg explicit timeout and cancellation behavior. Isolated review runs review steps in tool-less contexts while solver steps use the shared workspace. Fail-safe application applies no patch when the workflow is cancelled or fails validation. Validated routing verifies workflow definitions, model bindings, fallback behavior, and model availability before execution begins.

Where the Preview Stops

HydraFusion remains an active research effort. Results, models, workflows, availability, names, and product behavior may change as GitHub learns from the preview.

Feedback goes through /feedback in Copilot CLI or the GitHub Community discussion. The company believes the next real gain in coding agents will come from combining frontier intelligence with runtime orchestration. HydraFusion is their first bet on that idea.