GitHub Copilot Fusion GA: Half the Cost Per Task, Billed Per Model

GitHub made Copilot Fusion generally available at Universe on October 28. You stop picking a model. You pick an effort level, and Fusion picks the models and the workflow for each task. GitHub's headline: near-top coding scores at about half the cost per task. The part that decides your bill is in the billing line: Fusion charges you for every model it calls.

What actually happened

Per the GitHub changelog:

• Three effort levels. Standard for everyday coding, High for harder tasks, Max for the most complex work.
• Three execution patterns. Single sends the task to one model. Cascade lets a cheaper model draft, then a quality check accepts it or escalates to a stronger model. Critique has one model draft, a second review, and the first revise.
• Billing. Each model Fusion uses is billed at its standard rate. Paid plans get 30% off model costs for a limited time. GitHub does not give an end date.
• Admin controls. Fusion only uses models your enterprise and org policies allow. Admins can turn off a model for Fusion alone and set Fusion at any effort level as the default.
• Where it runs. All Copilot plans, in VS Code, the Copilot app, Copilot CLI and JetBrains IDEs. HydraFusion preview users move over automatically.

Artificial Analysis ran the benchmarks. Fusion Max (via Copilot CLI) scored two points below Claude Code with Opus 5.5 Max on its Coding Agent Index. It cost $5.59 per task against $11.42 and used 7.2M tokens per task, about half. The model pool is Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6 Sol. Fusion led on Terminal-Bench 4.0 (71) but trailed on FrontierCode (51 vs. 56). GitHub's own "40% less time" claim was not measured by Artificial Analysis.

Why Copilot Fusion matters for your business

We wrote about HydraFusion in September: cost dropped everywhere, quality did not. The GA numbers are better. A two-point gap at half the price is a trade most small teams should take for routine work.

But routing moves the cost decision from you to GitHub. Critique and cascade runs call more than one model per task. Your bill is per model, not per task, and the 30% discount will end.

Set the default effort to Standard. Make developers opt into Max for a reason. Max is the tier the benchmarks used.

Fence the model list. Disable models for Fusion that you have not cleared for your code or your data terms.

Price it on your repos. Run a week of real tickets on Fusion and on your current model. Compare cost per merged PR, not cost per request.

Plan for the discount ending. Budget at full rate now.

Sources: GitHub Changelog: Route and orchestrate models with GitHub Copilot Fusion, Artificial Analysis: GitHub Copilot Fusion, VentureBeat.