GitHub HydraFusion Outperforms Claude Opus 5 at Lower Cost

GitHub has introduced Project HydraFusion, a new runtime multi-model orchestrator available as a research preview for Copilot users. Instead of relying on a single large language model, the system dynamically routes developer queries across multiple AI providers. It automatically decides whether to draft, critique, revise, or escalate a task on the fly. This approach aims to optimize the balance between performance and cost by treating model selection as a real-time routing problem.

In benchmark evaluations against Claude Opus 5 and GPT-5.6 Sol, HydraFusion demonstrated significant cost savings with minimal impact on quality. On TerminalBench 2.1, the system achieved 4.9 more quality points than Claude Opus 5 while reducing estimated costs by 67 percent. On CheckpointBench, GitHub's internal replayable benchmark, it trailed Claude Opus 5 by just 0.1 quality points while cutting costs by 65 percent. For the repository-level tasks in DeepSWE, HydraFusion finished 1.5 quality points behind Claude Opus 5 but still lowered expenses by 36 percent.

To achieve these results, HydraFusion employs three distinct execution patterns based on task requirements: Single, Cascade, and Critique. The Single pattern routes a query directly to one model. The Cascade pattern uses an efficient model to draft a solution, escalating to a stronger model only if a quality gate fails. The Critique pattern uses one model to draft a solution and an independent, read-only critic from a different model family to review and suggest revisions. GitHub tuned these routing policies using a beam search over per-capability scores across all three benchmarks.

The orchestrator is currently available to all Copilot subscribers through the /experimental command in the Copilot CLI, and can be activated by typing /model and selecting HydraFusion. Users are billed at the standard token rates of the underlying models consumed, with no additional orchestration surcharge. While the system is currently optimized for first-turn, single-prompt tasks, GitHub plans to support multi-turn sessions in future updates. One current trade-off is that HydraFusion hides intermediate drafts until a final result is ready, which reduces real-time visibility during execution.