On October 14, 2026, GitHub updated HydraFusion, the system that distributes tasks across models within GitHub Copilot, and added HydraFusion Max for more demanding work. Both options are available in VS Code, the GitHub Copilot app and Copilot CLI, but remain in research preview. The practical change is a choice of processing depth, not a promise that every answer will improve.
The distinction starts with what users select. HydraFusion is not a new model itself: it chooses an execution pattern, models and reasoning effort for each step, then returns one final answer. One model may draft a solution and another review it. A straightforward task can also use just one model, without an unnecessary review pass.
What changes under the hood
GitHub says the model pool now includes Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. The actual combination depends on the workflow, so the list does not mean every request uses all of them. Reasoning effort also changes: medium for drafting and high for review, rather than relying on each model’s default setting.
The review patterns change too. Cascade can add a judging step that examines the initial solution and requests a repair only when necessary. Critique relies more on tools that inspect the actual workspace, less on written summaries, and includes parallel critic models. For a developer, that distinguishes reading a description of a change from examining the work the model actually performed.
Max uses the same architecture but targets higher effort where review depth matters. The documentation also describes extra-high reasoning effort. It is worth testing on a defined task with clear success criteria, such as a complex bug fix, and comparing solution quality, time and consumption. More reasoning steps do not replace running tests or human code review.
Who gets access and what to check first
The options are offered on Copilot Pro, Pro+, Business and Enterprise plans with experimental or preview features enabled, subject to organizational model policies. An eligible subscription alone does not override an administrator’s restriction. Model selection also remains with the orchestrator, using models permitted and available to the user, rather than becoming a manual choice for each step.
According to the documentation, billing covers every participating model at its standard rate, with no separate HydraFusion surcharge. The auto model selection discount does not apply, and a multi-model task may consume more AI credits than a single-model task. This describes the billing mechanism, not a subscription price change or a promise of fixed savings on every task.
Before team adoption, check permissions, monitor consumption and start in a test environment. GitHub states that the preview has no service-level agreement and is not intended for production workloads. The update broadens options for reviewing coding solutions, while leaving a simple decision with the user: does the extra review improve this particular task enough to justify the time and resources?