GitHub's HydraFusion picks a single, cascade or critique workflow per request and cuts TerminalBench cost 67% while gaining points

GitHub Blog·high signalXBlueskyLinkedInCopy link

GitHub published Project HydraFusion on 2026-09-04, a Copilot research preview that treats workflow selection as an optimization problem and routes each request to the least complex of three patterns: one model solving directly, a cheap model drafting with escalation to a stronger one, or a draft reviewed by a critic from a different model family before revision. Measured against Claude Opus 5 and GPT-5.6 Sol baselines, it reports 67% lower cost with +4.9 points on TerminalBench 2.1, 65% lower cost at -0.1 points on CheckpointBench, and 36% lower cost at -1.5 points on DeepSWE. The cross-family critique step is the part worth stealing: it is the reviewer coming from a different vendor that does the work, not the extra pass.