GitHub launches HydraFusion for task-specific model workflows
GitHub has released Project HydraFusion as a research preview in GitHub Copilot CLI. Instead of selecting one model for an entire request, the system creates an execution plan for each coding task and can use models from multiple providers. It is available to users on all GitHub Copilot plans. Developers can enable it with /update, /experimental on, and /model, then select HydraFusion (Research Preview). There are no open weights or self-hosted option, and usage is billed according to the standard token rates of the models involved.
HydraFusion currently chooses among three patterns. Single sends the task to one model. Cascade lets an efficient model draft a solution and escalates to a stronger model if a quality gate rejects it. Critique has a model produce a draft, an independent tool-less reviewer from another model family assess it, and the drafting model revise once. The runtime records cost, latency and outcomes for each stage, while cancelled or invalid workflows apply no patch.
In GitHub’s tests, HydraFusion achieved 4.9 more quality points than Claude Opus 5 on TerminalBench 2.1 while costing an estimated 67% less. On DeepSWE and CheckpointBench, it was 1.5 and 0.1 points lower, with estimated savings of 36% and 65%, respectively.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.