GitHub brings HydraFusion’s multi-model coding workflows into VS Code

On September 30, 2026, GitHub expanded its HydraFusion research preview from Copilot CLI to Visual Studio Code 1.140 or later and the GitHub Copilot app. HydraFusion is presented in the model picker but is not itself a model. For each turn it can choose a Single workflow, a Cascade that starts with an efficient model and escalates if a quality gate rejects the result, or a Critique workflow where a read-only critic from a different model family reviews a draft before one revision. GitHub also added more visible workflow status and progress reporting. Business and Enterprise administrators must allow preview features.

Most coding assistants expose model choice as the main intelligence control. HydraFusion moves that control one layer upward: the platform decides whether a task needs one model, conditional escalation or a second model acting as critic. Expanding this from a CLI experiment into VS Code makes compound-model execution available in a mainstream coding surface without developers manually coordinating agents or model calls. It also changes how cost should be evaluated: GitHub bills the underlying model tokens, so the useful metric is completed-task cost and quality rather than the sticker price of one selected model. The caveat is substantial: HydraFusion remains a research preview, its model pool is not fully exposed, and GitHub's benchmark results are vendor-produced rather than independent evidence.

Copilot Auto selects one model for a request. HydraFusion can instead select among Single, Cascade and Critique patterns. Cascade uses an efficient first attempt plus a quality gate that can escalate to a stronger model. Critique uses a separate read-only model family to review the draft before the drafting model revises once.

HydraFusion now appears in the Copilot Chat model picker in VS Code 1.140+ and can be enabled in the Copilot app. That removes the earlier Copilot CLI-only boundary. Organizations using Business or Enterprise plans must permit preview features before users can select it.

GitHub's earlier controlled offline evaluations reported estimated workflow-cost reductions of 36% to 67% against its Claude Opus 5 baseline across three coding-agent benchmarks. Quality was 4.9 percentage points higher on TerminalBench 2.1, 1.5 points lower on DeepSWE and 0.1 point lower on CheckpointBench. Those results describe GitHub's own test setup and should not be assumed to transfer to a particular repository or workload.

The practical distinction from Auto is architectural. Auto decides which single model should receive a prompt. HydraFusion decides how many model roles the turn needs and how they interact. If this approach graduates from preview, developers may increasingly choose an optimization policy while the coding platform manages model composition underneath.