Copilot adds HydraFusion Max for more demanding coding tasks

Some website bugs need a second look: the first edit fixes one screen but breaks another. GitHub’s October 14, 2026 HydraFusion update changes how Copilot combines drafting and review, and adds a Max option for more demanding work. Both remain research previews. Dated release.

What the extra pass does

HydraFusion selects an execution pattern and the models that run it. A simple task can use one model. Cascade can have a judge inspect an initial solution before escalation, while Critique combines independent feedback and revises the draft. The update also puts more emphasis on inspecting the actual workspace. How the patterns work.

Max applies higher reasoning effort where depth matters. That gives you a new option to evaluate, rather than a measured guarantee that every fix improves.

Try an existing bug with a clear acceptance check. For example, a newsletter form should show a useful error for an invalid address and stay usable after correction. Keep that behavior fixed when comparing approaches. Review the changed files, the check result, elapsed time and reported usage. This is an illustrative comparison; we have not run it.

Review the bill and the files

GitHub bills each participating model at its standard rate. There is no separate HydraFusion fee, but the Auto discount does not apply and extra passes can consume more credits. In Copilot CLI, /usage shows usage by model. Billing and usage, current model rates.

Another useful limit: discarding a draft does not automatically undo edits it already made. Inspect the workspace before committing, even when the final answer looks tidy.

The release lists Pro, Pro+, Business and Enterprise access across VS Code, the Copilot app and CLI, with preview features enabled and model policies applied. GitHub’s documentation says this research preview has no SLA and is not intended for production workloads. Use an isolated evaluation task first.

Our Gemini API and Copilot access story covers choosing a particular model. Here, the decision is whether the additional drafting and review steps help your bounded task enough to justify their time and usage.