GitHub has introduced Project HydraFusion, an advanced research preview for GitHub Copilot designed to deliver frontier-level coding intelligence through runtime model orchestration. Building on previous automatic model selection capabilities, HydraFusion treats workflow execution as an optimisation challenge, dynamically building full execution plans using models from multiple providers to handle core developer tasks.
The system evaluates incoming prompts using explicit capability signals tailored for complex operations, including multi-step reasoning, automated code generation, structured debugging, and advanced tool use. Rather than relying on a single static model, HydraFusion routes requests across three distinct runtime execution patterns depending on the task's complexity and context:
- Single: One selected model executes directly when it has sufficient capability to solve the task independently, optimising for speed and low latency.
- Cascade: An efficient model generates an initial solution draft, which a quality gate evaluates. If the output satisfies requirements, it is accepted; otherwise, the task escalates to a stronger, more capable model.
- Critique: A drafting model produces an initial solution artefact, which is then assessed by an independent, read-only critic model drawn from a separate model family without tool execution access (mirroring the Rubber Duck review pattern). The original drafting model then performs a single structured revision based on this review.
To ensure robust and production-grade execution, HydraFusion’s architecture is anchored by five fundamental operating principles. These include complete accounting to track token cost and usage across every workflow leg (drafting, critique, revision, escalation, retry, and fallback), bounded execution enforcing strict timeouts and cancellation handles, isolated review steps that prevent modifier actions within a tool-less environment, fail-safe application routines that reject patches if validation fails or execution is cancelled, and validated routing to pre-check model availability and bindings before runtime kicks off.
In controlled offline evaluations across three agentic coding benchmarks, HydraFusion’s selective runtime workflows matched or exceeded baseline quality metrics while substantially reducing estimated costs. Notably, on TerminalBench 2.1, it delivered a 4.9 percentage point improvement in verified task quality while achieving a 67% reduction in estimated cost compared to Claude Opus 5.
On CheckpointBench, an internal multi-turn benchmark curated from real, replayable GitHub Copilot agentic coding sessions anchored to specific public repositories and immutable commits, the project achieved a mean session score virtually tied with the Claude Opus 5 reference baseline, registering a minor 0.1 percentage point difference while lowering estimated workflow costs by 65%.
Project HydraFusion is currently available as a research preview to users across all GitHub Copilot tiers via the /experimental configuration inside the GitHub Copilot CLI. Developers can engage the feature by updating their CLI environment, executing /experimental on, and selecting HydraFusion from the /model selection interface, with usage billed according to the standard token rates of the underlying models invoked during execution.