What changed
On October 28, 2026, GitHub renamed and generally released its HydraFusion multi-model orchestration as Copilot Fusion across VS Code, Copilot CLI, Copilot app and JetBrains. It selects Single, Cascade or Critique execution per task and offers Standard, High and Max effort tiers. Enterprise administrators can restrict the eligible models and set a default. Billing charges each underlying model at its standard rate, with a limited-time 30% discount on model costs for paid plans. In a separate experimental preview, Copilot Fusion Standard can route to local models: managed Windows ML setup for MAI Code 1.1 Flash on supported Windows devices, or manually configured OpenAI-compatible local servers on Windows, macOS and Linux through Copilot CLI. Routing decisions still happen in the cloud, so local inference does not make sessions offline.
Why it matters
The original October 1 dossier described HydraFusion as a research preview for mainstream editors. General availability changes procurement and operational planning: teams can now select an effort policy instead of choosing a single model, restrict the model pool under enterprise governance and estimate costs from all participating model calls rather than a single model price. GitHub reports Fusion Max scored 60 on Artificial Analysis Coding Agent Index, within two points of Claude Code Opus 5.5 Max, at roughly half average API cost per task and 40% less time; these are GitHub-cited benchmark comparisons, not independently reproduced measurements of a team's workload. The local-model preview may reduce some cloud inference work but still requires cloud routing and suitable local hardware. Admins should verify which models, data policies and spend limits apply to their organisation.
October 28: research preview becomes Copilot Fusion GA
Fusion replaces the HydraFusion research-preview name and is available on all Copilot plans in VS Code, Copilot app, CLI and JetBrains. Standard favours versatile everyday models, High more capable models for harder tasks, and Max the strongest models for demanding work. The system can answer with one model, escalate from an efficient model after a quality gate, or have a read-only critic review and request a revision.
Costs and enterprise model controls are explicit
GitHub bills each underlying model at its normal rate, not a flat fee per Fusion turn; paid Copilot users temporarily receive 30% off model costs through Fusion, with no published end date in the announcement. Administrators can enable or disable particular models for Fusion independently of their general availability, and set Fusion as a default through managed settings. Compare total task cost, latency and review quality rather than just advertised model multipliers.
Experimental local routing is not an offline Copilot
On supported Windows devices, Windows ML can download and configure MAI Code 1.1 Flash. Copilot CLI also supports manually configured MAI Code 1.1 Flash or Qwen3.8 Flash Next through an OpenAI-compatible local server on Windows, macOS or Linux. The Copilot app's intelligent local routing currently supports managed Windows ML only. Local-model eligibility is separate from generic BYOK discovery. Routing decisions still occur in GitHub's cloud and local models can have different safety protections.
How to interpret the benchmark and rollout
GitHub cites a Fusion Max score of 60 on Artificial Analysis Coding Agent Index, within two points of Claude Code Opus 5.5 Max, at about half average API cost and 40% less task time. Those numbers describe the reported benchmark setup and do not establish equivalent savings for all repositories. GA applies to the core Fusion orchestration; local-model routing remains an opt-in experimental preview.