GitHub HydraFusion Routes Coding Work Across Models
GitHub’s experimental coding agent selects single, cascade or critique workflows to trade model cost against verified task quality.
One request, several possible workflows
GitHub has opened a research preview of Project HydraFusion, a Copilot runtime that decides both which AI models should handle a coding task and how they should collaborate. Instead of sending every request to one chosen model, HydraFusion can use a single-model response, let a cheaper model attempt the work before escalating, or have a second model critique a draft before revision.
The preview is available across Copilot plans through the experimental menu in GitHub Copilot CLI. Users select HydraFusion as if it were an ordinary model, while GitHub’s runtime constructs the workflow. Charges reflect the tokens consumed by every participating model at its standard rate, including drafting, review, revision and escalation.
GitHub says HydraFusion improved verified task quality by 4.9 percentage points on TerminalBench 2.1 while cutting estimated cost by 67% relative to Claude Opus 5. Its policies were refined against TerminalBench, DeepSWE and a GitHub dataset derived from real Copilot sessions. The system applies timeouts, validates model bindings and runs critics without tools so that reviewers cannot modify the shared repository. Failed or cancelled workflows are designed to apply no patch.
The evidence remains bounded: these are controlled offline evaluations, and GitHub recommends starting with substantial, single-prompt tasks. Multi-turn sessions, latency and behavior on messy production repositories still require broader testing.
Why it matters
HydraFusion turns model selection into an infrastructure decision rather than a choice repeatedly imposed on developers. If its routing survives real workloads, coding-agent vendors could combine inexpensive drafts with selective frontier-model calls instead of tying product quality and margins to one provider. That also gives GitHub greater control over the interface between model suppliers and Copilot customers—a strategically valuable position as capable models become more interchangeable.