GitHub announced on October 28, 2026 that Copilot Fusion, a model-picker option that no longer locks a task to a single model but splits each request across several models depending on the chosen effort level, has reached general availability. The stated goal is to balance performance and cost without leaving that choice to the developer on every single prompt.
The problem Fusion is meant to solve
Picking a powerful model for a hard reasoning task, then forgetting to switch back to a lighter model for routine work, is a well-documented source of silent overspend among teams using coding assistants daily. Fusion moves that decision from the developer to the tool: for every prompt, the system analyzes the task and picks the model combination it judges best suited.
Three orchestration strategies
Three workflows are available depending on the effort level chosen in the model picker (Standard, High, Max):
| Workflow | How it works | Typical use case |
|---|---|---|
| Single | One model handles the request end to end | Simple, well-defined task |
| Cascade | An efficient model drafts an answer; a quality check accepts it or escalates to a stronger model | Task of uncertain difficulty |
| Critique | One model drafts, a second model reviews, the first revises | Sensitive reasoning or refactoring task |
GitHub does not publish the exact list of models Fusion orchestrates: the feature combines the multi-model approach from its internal HydraFusion project with the automatic model selection already built into Copilot, and only draws on models enabled under an organization’s or enterprise’s model policy.
Results benchmarked against Claude Code
On the Artificial Analysis Coding Agent Index, GitHub places Fusion Max at a score of 60, within two points of Claude Code running Opus 5.5 Max, used as the market benchmark. The headline figure is mostly about cost: GitHub claims this result at roughly half the average API cost per task, with about half the tokens consumed and close to 40% less processing time per task.
Fusion moves the tradeoff between power and cost from the developer to the tool, request by request.
Rollout, billing and admin controls
Fusion is available on every paid Copilot plan, in Visual Studio Code, the Copilot app recently updated with computer use, the GitHub Copilot CLI and JetBrains IDEs; other clients are coming later. Accounts that were part of the research preview under auto, HydraFusion or HydraFusion Max are automatically moved to the matching Fusion effort level.
On billing, every model Fusion calls is charged at that model’s standard rate, but GitHub is applying, for a limited time, a 30% discount on model costs incurred through Fusion for accounts on a paid Copilot plan. Organization and enterprise admins can enable or disable specific models for Fusion, and set Fusion, at a given effort level, as the default model through enterprise-managed settings.
What this means for a web development team
For a team already managing the cost of agentic fixes running in production, Fusion offers a reasonable default without fine-tuning model by model. It remains worth setting the default effort level at the organization level rather than leaving it to individual choice, to avoid a routine task systematically running at Max.
Key takeaways
Copilot Fusion generalizes a multi-model orchestration approach already tested in preview, with three clearly defined strategies and a promise of lower cost at comparable quality. Availability across every paid plan and the main IDEs makes it worth testing quickly, while keeping an eye on per-model billing rather than the headline Fusion rate alone.
I stay cautious about vendor-published benchmark comparisons: a two-point gap on a composite index says little about behavior on a real project. What strikes me as more solid is the conditional-escalation principle itself, which matches what I already do manually when moving from a lightweight model to a more capable one only after the first pass fails. Automating that reflex makes sense, as long as you keep control over the default effort level. — Simon Janvier
Further reading: the full breakdown of how Copilot Fusion works and its billing terms is available in the official GitHub changelog entry.
