GitHub Copilot Fusion: Stop Picking a Model, Start Picking Who Does Each Step
GitHub made Copilot Fusion generally available at Universe on 28 October 2026. It is a runtime router and orchestrator: you pick an effort level, it decides which model — or which combination of models — handles each step. GitHub says it grew out of Project HydraFusion's roughly 60,000 routed sessions and Copilot's auto model selection, and that it only draws from models your enterprise and organisation policies already allow
Official Links
Features
- ✓ Splits a coding task into steps and picks the model and execution pattern per step, so one model can draft while another reviews
- ✓ Nothing to name in the model picker: choosing Copilot Fusion means choosing a runtime routing system, then setting the effort level yourself
- ✓ Respects enterprise model policies: admins can see which models Fusion draws from and disable a single model for Fusion without disabling it for everyone
- ✓ Shows progress per step and which model is running; intermediate drafts may be revised or discarded and only the final response is returned
- ✓ The stated motivation: in September 2026 paying Copilot developers used 47 different models, and no single model handled more than a quarter of their requests
How to Use
- Confirm your Copilot plan includes Fusion, then check which models are permitted in your enterprise or organisation model policies
- Pick GitHub Copilot Fusion in the model picker and set the effort level: Standard for routine edits, higher for complex refactors or cross-file debugging
- Run a familiar task you have done before and compare output, cost and time against your previous habit of hand-picking a model
- For teams, have an admin review the Fusion model pool first, then decide which models stay available to Fusion and which remain manual-only
FAQ
What problem does Copilot Fusion actually solve?
Per GitHub's changelog, instead of picking a model for every task you select Fusion and let Copilot route and orchestrate models at runtime to balance performance and cost. GitHub's own motivation data: in September 2026 paying Copilot developers used 47 different models and no model exceeded a quarter of requests, which means there is no longer a single best model for everyone. Docs at https://docs.github.com/copilot/concepts/models/copilot-fusion and the changelog at https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion
How is it different from the earlier auto model selection?
GitHub describes Fusion as bringing together Project HydraFusion's multi-model orchestration and Copilot auto model selection's fine-tuned selection logic. The distinction GitHub itself draws is that Auto picks the optimal model, while HydraFusion picks the optimal workflow, possibly using several models. Fusion is the GA product combining both, and GitHub has said it expects the experiences to converge over time, so team guidance should cover both for now
Are there public numbers on quality and cost?
Artificial Analysis benchmarked the Max setting in GitHub Copilot CLI on 28 October 2026 using its Coding Agent Index v2.0: about 2 points below Opus 5.5 in Claude Code, at roughly half the cost per task and about 40% less time, with 71 on Terminal-Bench 4.0 (a lead), 51 on FrontierCode (third) and 57 on SWE-Atlas-QnA-AA. Those are third-party public results at https://artificialanalysis.ai/articles/github-copilot-fusion-orchestrates-anthropic-and-openai-models-to-balance-performance-and-cost — GitHub cited similar figures at Universe, but your own workloads still need testing
How is it governed in an enterprise?
The official docs state Fusion only uses models enabled in your enterprise and organisation model policies. Administrators can view the candidate pool, independently set each model's effective policy state, and disable a model for Fusion alone without removing it for everyone. Governance stays in the existing model policy surface rather than a parallel permission system
Which teams should adopt it first?
Teams already switching between several models, or arguing over which one should take a task, get the clearest benefit: the choice moves into the system and humans keep direction and effort level. If you run one stable model in a fixed pipeline and value predictability above all, manual selection stays more transparent — pilot Fusion on a small set and A/B it on the same tasks before rolling out