Blog / Developer tools

What Is GitHub Copilot HydraFusion? The New Max Option Explained

2026-10-18cdnping editors4 min read

HydraFusion is a model orchestration option in GitHub Copilot. Behind the scenes it coordinates several models and merges their work into a single answer, so you no longer have to pick a model for every task. On October 14, 2026, GitHub updated HydraFusion in its changelog and added HydraFusion Max. As of October 2026 it is still a research preview.

Below are the questions readers ask most often. Everything here comes from that changelog. Anything it does not mention (pricing, general availability date) is not guessed at in this article.

#What is HydraFusion?

In short, it is not a new model. It is a dispatch pipeline: one request passes through several models and several steps, and one result comes out. The changelog says the model pool includes models from multiple vendors, and which ones are used depends on the orchestration workflow you choose.

You can use it in three places: VS Code, the GitHub Copilot app, and Copilot CLI.

#What changed in the October 14 update?

The changelog lists a few changes:

• Reasoning effort is no longer a single default for every step. The first draft uses medium effort, and the review step uses high effort.
• Cascade workflow: an initial model answers first and escalates to a stronger model when needed. It now has a judging step that checks the first draft and asks for fixes only when necessary.
• Critique workflow: reviews use tools to inspect the actual workspace instead of relying mainly on a text summary, and they use multiple parallel reviewer models.

#How does HydraFusion Max differ from the standard version?

Max uses the same architecture and workflow techniques but is tuned for capability: it applies higher reasoning effort at the steps that need depth, and it targets tasks that call for the deepest reasoning and the most careful analysis. It still considers token efficiency through orchestration, but GitHub positions it as trading some of the token savings for higher capability.

The changelog does not give Max's specific reasoning-effort levels, and it does not give any figure for how many tokens are saved.

#Who can use it, and how do you turn it on?

• Plans: Copilot Pro, Pro+, Business, and Enterprise.
• You need to enable experimental or preview features.
• It is subject to your organization's model policy. If you are on an enterprise or team account and the option is missing, ask your admin whether it is allowed.
• The changelog says nothing about pricing or how usage is counted against your quota. For billing, check GitHub's official documentation.

#Checklist for trying it

1. Confirm your plan is in the list above. On an organization account, check the admin's model policy first.
2. Enable preview features in VS Code, the Copilot app, or the CLI (see the official docs for the exact menu location).
3. Look in the model picker of your client to see whether HydraFusion or HydraFusion Max appears (see the official docs for where it lives).
4. Take a real task you know well and run it three times: with the model you usually pick by hand, with HydraFusion, and with Max.
5. Record three things: whether the result is usable as is, how long you waited, and how much quota it used. Then decide which one to default to.

#Picking models yourself vs. letting orchestration do it

The upside of picking a model by hand is control and reproducibility: when something goes wrong, you know which model did it. The upside of orchestration is that you skip the decision. Hard problems escalate automatically, and easy ones do not waste capacity. The cost is that you cannot see the intermediate steps. When something goes wrong, it is hard to tell which model or which step caused it, and reproducing the problem is harder. That matters even more during a preview.

This idea of routing by task difficulty is not unique to Copilot. If you want to build it yourself, see how to do multi-model routing. For the differences between ways of accessing models, see official APIs, cloud-hosted models, and aggregation gateways compared.

AGI Plan is a commercial partner of cdnping. Its Prism scheduling layer also routes requests to different capability groups based on task difficulty. You use it by changing one endpoint address in tools such as Claude Code or Codex; for setup, see custom API and endpoint configuration. For the models and regions currently available, go by agiplan.dev/models.

#What to watch for with preview features

A preview feature's model pool, behavior, and limits can all change. Copilot has also made some recent model retirements, so read this alongside Copilot retiring three models on November 16. To learn about changes to official pages as soon as they happen, use cdnping's site changes page, or read about how to track price and model changes. For model availability across inference providers, see model monitoring.

#Sources

• GitHub Changelog: HydraFusion updates and new Max option (2026-10-14)