HydraFusion Max vs HydraFusion: What Changed in GitHub Copilot

GitHub introduced HydraFusion Max on October 14, 2026, alongside changes to its existing HydraFusion orchestrator in Copilot. These are not two new standalone AI models. Both coordinate multiple models and reasoning steps behind a single response; Max is tuned for especially demanding work, while the standard option balances capability, latency and AI credit usage. Both remain in research preview.

HydraFusion vs HydraFusion Max: the decision

GitHub describes HydraFusion as a runtime orchestrator rather than a model with its own fixed context window. It chooses an execution pattern and model combination for each request. Max uses the same architecture but can apply extra-high reasoning effort where deeper scrutiny matters. Neither option guarantees a better answer on every task, and both can change during the preview.

What changed in the October 14 update

• The available model pool now includes Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna; the exact models used vary by task.
• Reasoning effort is now tuned by step, with medium effort for drafting and high effort for reviewing.
• Cascade can use a judging step to accept a first attempt or request a repair instead of always escalating.
• Critique now uses tools to inspect the workspace more directly and parallel critic models to review work.
• HydraFusion Max adds a deeper reasoning configuration for difficult tasks without requiring manual model switching.

How the orchestration patterns work

Single uses one model. Cascade drafts with one model, then judges whether the work needs a stronger pass. Critique drafts, asks multiple critics to review in parallel, and revises once. These are execution patterns selected by the system, not switches that users must set per prompt. For a simple change, Single may be quicker; for a difficult debugging task, a review pattern may justify its overhead.

Availability, preview limitations and billing

Both options are available in Copilot CLI, Visual Studio Code and the GitHub Copilot app for eligible Pro, Pro+, Business and Enterprise users with preview features enabled, subject to organization model policies. GitHub's documentation says the research preview has no service-level agreement and is not intended for production workloads. Organizations should test in non-production environments and confirm their administrators allow the feature.

GitHub charges for each model used at its normal rate; there is no separate HydraFusion charge, and the auto model-selection discount does not apply. Multiple passes can therefore use more AI credits than one-model work. GitHub does not publish a universal cost or speed advantage for Max. Compare actual credits per accepted change rather than assuming that more reasoning is always economical.

A controlled evaluation checklist

1. Choose 10 to 20 representative repository tasks, including simple edits, multi-file refactors and hard debugging cases.
2. Use the same code snapshot, acceptance tests, task instructions and permissions for Auto, HydraFusion and Max.
3. Track successful tests, correctness after human review, elapsed time, AI credits and rework.
4. Review every changed file and tool action: a discarded draft can leave workspace edits behind, according to GitHub.
5. Select the least expensive option that consistently meets the quality bar, and keep preview workflows away from unattended production changes.

How to enable it

In Copilot CLI, update to a current version, enable experimental features and select HydraFusion or HydraFusion Max from the model picker. In Visual Studio Code, use version 1.140 or later and enable the chat.copilot.hydraFusion.enabled setting. In the Copilot app, enable HydraFusion under Experimental settings. If neither option appears, check the plan, preview policy and model permissions before troubleshooting the client.