GitHub added a Max tier to its HydraFusion Copilot orchestration

A multi-model orchestrator that always picked the balanced path left nothing for the task where a developer wanted the deepest answer the system could produce. On 14 October 2026, GitHub added that option to HydraFusion, its Copilot feature that chooses between a single model, a cascade or a critique for each request, and called the new mode HydraFusion Max.

Both modes remain in research preview and run across VS Code, the GitHub Copilot app and the Copilot CLI. The product page describes HydraFusion as a system that "orchestrates multiple models to produce a single response", and the Max option is "built on the same architecture and uses the same workflow techniques" with "higher reasoning efforts where depth and thoroughness matter most".

What the two options are for

GitHub positions HydraFusion as a balanced mode for "agentic tasks that require a balance of speed, cost, and performance". HydraFusion Max is for "tasks requiring the deepest possible reasoning and most thorough analysis", with orchestration still used to hold token spend down. The post does not state how much Max costs in tokens against the standard mode, so a reader has to try both on their own work to see the gap.

The model pool and the effort split

HydraFusion and HydraFusion Max draw from a pool that now includes Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, and GitHub says the model used for a given task varies with the orchestration workflow. The reasoning effort setting changed with it: HydraFusion used to run each model at its default effort, and now applies medium effort to initial drafts and high effort to review. GitHub frames the split as "keeping drafting efficient while dedicating deeper reasoning to checking and improving the work".

Two workflows got a step added

Cascade used to start with one model and escalate to a stronger one on need. The new release puts a judging step between those two, which inspects the first attempt and asks for a repair only when the judge rejects it. Critique used to rely on text summaries of the workspace; it now reads the workspace through tools directly and runs several critic models in parallel on the same work.

Neither change is a new idea in the agent-engineering literature. What is new is that a developer who uses Copilot can pick a workflow that applies them, next to the other models in the picker, and watch which path the system took on each turn.

Who can turn it on

HydraFusion and HydraFusion Max are available on Copilot Pro, Pro+, Business and Enterprise, with experimental or preview features enabled, and are subject to the organisation's own model policies. Business and Enterprise administrators have to switch preview features on for the organisation first. GitHub keeps the research preview label, which means the behaviour can change without notice.

A team deciding between the two modes will want to measure them on the same tasks before standardising on either. The owner of a budget cares about which share of requests Max actually earns, and the owner of a review queue cares about whether the deeper review path cuts the second round-trip that lands on a human.

Source

• GitHub Changelog, HydraFusion updates and new Max option