GitHub's HydraFusion Max doesn't pick a model for your task. It plans who drafts, who judges and who repairs

The first generation of model routing asked one question per request: which model should answer this? GitHub’s October 14 update to HydraFusion, its multi-model orchestration for Copilot, asks a different one: which model should do which part of the work, and how hard should each think?

HydraFusion now draws on a pool of Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, and GitHub added a Max option tuned for the most demanding tasks, with higher reasoning effort where depth matters, while orchestration still weighs token efficiency. Both remain in research preview, for Copilot Pro, Pro+, Business and Enterprise plans with experimental or preview features enabled, and subject to the organization’s model policies.

What changed in the workflows

Two orchestration patterns were reworked, and they show where the design is going.

1. Draft at medium reasoning effort
2. A judge inspects the initial work
3. Repair requested only if needed

Spend on producingSpend on checking

Cascade used to let an initial model try a task and escalate to a stronger one when needed. It now adds a judging step that inspects the initial work and asks for a repair only if necessary. Critique used to depend heavily on text summaries of the work; it now uses tools that inspect the actual workspace and runs several critic models in parallel. And reasoning effort is no longer each model’s default: HydraFusion sets medium effort for initial drafts and high effort for review.

Each of those choices moves spending from producing to checking. Drafts get cheaper; judgment gets more attention. It is an execution strategy, not a model choice.

Why critics that read the workspace matter

A reviewer that only reads a summary of a change reviews the author’s account of the work, not the work. Critics that inspect files, run tools and see the repository can catch what the summary left out. That is the same distinction between a claim and evidence that runs through good human code review, and it is a real improvement in a system that reviews itself.

It is still the system reviewing itself. Parallel critics drawn from the same model pool are a stronger check than one critic, but they are not independent in the sense a deterministic test or a human owner is. Treat HydraFusion’s internal review as quality improvement before the pull request, not as a substitute for the review after it.

How to evaluate it

More model calls per task can raise quality, cost or both. GitHub’s changelog does not give pricing or a cost comparison for the Max option, so the economics have to be measured on your own work:

1. Pick a fixed set of real tasks from your repositories, with known good outcomes, and run them through your current default, HydraFusion and Max.
2. Measure cost per accepted change, not cost per request. A run that costs three times as much and merges without rework can be cheaper than three cheap attempts and a frustrated reviewer. We made this case for inference priced per completed task.
3. Record time to a mergeable change, because multi-stage orchestration adds latency even when it saves review.
4. Decide where Max is allowed. A tier built for the hardest tasks will be used for easy ones if it is the default. Make it an explicit choice per repository or task type.

A preview of how model portfolios will run

HydraFusion is a coding feature, but the pattern generalizes. We argued that enterprises would run a portfolio of models with distinct roles and that OpenAI’s own guidance points toward routing by task. GitHub takes the next step: routing by stage within a task, with a budget for thinking at each one.

For platform teams, that changes what policy has to express. “Which models are approved” is no longer enough. The questions become which models may draft, which may judge, how much reasoning each stage may spend, and who sees the bill. Settle those while the feature is still a preview, and the first production rollout will not settle them for you.