GitHub's HydraFusion Cuts Copilot Cost, Trims Quality Twice

GitHub opened a research preview of Project HydraFusion on Sept. 4, according to the GitHub Blog post announcing it, a system that builds a different execution plan for each coding task inside GitHub Copilot CLI instead of sending every request to one model.

Anyone on any Copilot tier can turn it on today by typing /experimental on in the CLI, GitHub said, with usage billed at each underlying model's standard token rate rather than a flat preview price.

Three workflows, chosen per task

HydraFusion picks from three patterns for a given request: Single, where one model answers directly; Cascade, where a cheaper model drafts and a quality gate decides whether to escalate to a stronger one; and Critique, where a second, different model reviews the first model's draft and the original model revises once based on that feedback. GitHub credited five named engineers and scientists with the project, including Principal Software Engineer Andy Salerno, who said in the post that "the reasoning and task solving capability [of HydraFusion] is at or better than Opus," referring to Anthropic's Claude Opus 5.

Where the numbers hold up, and where they don't

GitHub's own offline benchmarks, independently reported by Olimpiu Pop at InfoQ on Sept. 13, show HydraFusion beating Claude Opus 5 outright on only one of the three benchmarks GitHub cited.

Benchmark	Cost vs. Opus 5	Quality vs. Opus 5
TerminalBench 2.1	67% lower	+4.9 points
DeepSWE	36% lower	-1.5 points
CheckpointBench	65% lower	-0.1 points

On TerminalBench 2.1, HydraFusion is both cheaper and more accurate. On DeepSWE and CheckpointBench, it costs 36% and 65% less than Opus 5 but trails it on quality, by 1.5 points and 0.1 points respectively. GitHub's post frames the project around the TerminalBench result and the Salerno quote; it does not lead with the two benchmarks where the quality line moved the other way.

What GitHub has not said

Neither the GitHub Blog post nor InfoQ's write-up names which models sit inside HydraFusion's orchestrated pool. Claude Opus 5 and OpenAI's GPT-5.6 Sol appear only as comparison baselines in the benchmark tables, not as confirmed members of the rotation, and GitHub has not published which model or models actually draft, critique or get escalated to inside a Cascade or Critique run. That leaves the mechanism behind the cost cuts, which model is doing the cheap first pass and which one gets called in when the quality gate trips, undocumented for now. GitHub also has not said when, or whether, HydraFusion moves from research preview to a supported, generally available feature, or what it will cost once the preview ends.

The preview follows GitHub's separate push to make Copilot's underlying AI usage cheaper without cutting task quality, and sits alongside VS Code 1.137's own scheduled agent automations, part of the same pattern of letting Copilot run tasks with less input from the person who filed them.