GitHub is testing a different answer to the model-selection problem in AI coding: do not make the developer choose one model at all.
Project HydraFusion, announced September 4, is now available as a research preview in GitHub Copilot CLI. The runtime builds an execution plan for each request and can choose among models from multiple providers. Depending on the task, it may use one model directly, start with a cheaper model and escalate if a quality gate fails, or have a separate model critique and revise the draft.
That is meaningfully different from Copilot's existing Auto model selection. Auto chooses a model for the task; HydraFusion can choose a multi-step workflow and use more than one model inside the same request. GitHub says the goal is to balance quality, cost and latency without requiring developers to manually orchestrate those trade-offs.
The preview is available through Copilot CLI's experimental mode across Copilot plans. Usage is metered according to the tokens consumed by the underlying models, at their standard rates. That means HydraFusion is not a flat-cost shortcut: a cascade or critique path can invoke several models, even if GitHub's routing is designed to use the least complex workflow expected to meet the quality bar.
GitHub reports lower estimated cost across its controlled offline tests and highlights a result that beat an evaluated Claude Opus 5 baseline. The benchmark table is more mixed on quality than that headline suggests. VentureBeat's independent read of the same table found that cost fell in all three reported benchmarks, while HydraFusion matched or exceeded the Opus 5 quality score in only one; the other two were slightly lower. The strongest launch claim therefore remains vendor-run evidence, not an independent benchmark of everyday Copilot use.
For Copilot, the broader significance is architectural. GitHub has spent 2026 adding more foundation models to the product; HydraFusion tests whether the next layer of differentiation is the orchestration policy that decides when to draft, escalate and critique. If it graduates from research preview, developers may increasingly select an execution strategy rather than a named model.
For now, teams should treat HydraFusion as experimental. It can be useful for evaluating compound-model coding workflows, but its behavior, model pool, latency and cost profile can change as GitHub tunes the preview. AiToolMap has also updated its prepublication GitHub Copilot review to reflect HydraFusion without changing the product's numeric rating: GitHub's own benchmark results are not used as independent rating inputs.