GitHub releases HydraFusion research preview for multi-model orchestration in Copilot CLI

GitHub shipped Project HydraFusion yesterday as a research preview, a system that selects not just a model but an entire workflow for every request in Copilot CLI. Instead of routing the prompt to a single pre-chosen model, the engine builds an execution plan in real time: it can draft with an efficient model, hand the draft to a critic from a different model family, or escalate to a stronger model when a quality gate rejects the first attempt. The models come from different providers, and the developer picks HydraFusion once, the same way they would pick any other model.

Three execution patterns

The system currently chooses among three patterns for each task. **Single** — one model solves the task directly, preserving speed. **Cascade** — an efficient model drafts a solution, a quality gate decides whether to accept it or escalate to a stronger model. **Critique** — one model drafts, a read-only critic from a different model family reviews the draft, and the drafting model revises once; the critique pattern follows the same principle as rubber-duck debugging. Each pattern balances quality and cost differently, and the engine picks the simplest flow expected to clear the quality bar.

Five engineering principles

GitHub built the runtime around five operational principles that matter for repository-level work: full accounting for every step — drafting, critique, revision, escalation, retry, and fallback; scoped execution with explicit timeouts and cancellation for each step; isolated critique where critics run without tool access and cannot modify the repository; safe application, so no edit is applied when the workflow is cancelled or fails validation; and verified routing, checking model links, fallback behavior, and availability before execution starts. Internally the system logs role, outcome, cost, latency, and diagnostics for every step; externally the developer sees a single coherent response and a permission-aware change set.

Benchmark results

The GitHub team evaluated fixed HydraFusion policies on three agentic coding benchmarks, using Claude Opus 5 and GPT-5.6 Sol as baselines, all at medium reasoning. The figures below are relative to Opus 5. On TerminalBench 2.1 the system gained 4.9 quality points at an estimated 67% lower cost. On DeepSWE and CheckpointBench it trails Opus 5 slightly while cutting cost by 36% and 65% respectively. CheckpointBench is GitHub's internal multi-turn suite, curated from real Copilot sessions and anchored to immutable public commits to enable repeatable runs.

Availability and billing

HydraFusion is available now on all Copilot plans, but only inside GitHub Copilot CLI. To enable: run `/update`, then `/experimental on`, then `/model` and select HydraFusion (Research Preview). There are no open weights and no self-hosted path. Billing is per token consumed by the models the workflow invokes, at each model's standard rate. Details are in the GitHub blog and community discussion #206492.