GitHub Copilot Fusion picks the right AI model for each coding task

What's new

GitHub has launched Copilot Fusion, a system that decides on its own which AI model should handle a coding task, so developers don't have to guess. According to the independent analytics firm Artificial Analysis, Fusion chooses between Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra and GPT-6 Sol.

There are three tiers: Fusion Standard for everyday coding, Fusion High for difficult tasks and Fusion Max for the most demanding tasks that require complex reasoning. Fusion is generally available on all Copilot plans in Visual Studio Code, the GitHub Copilot app, Copilot CLI and JetBrains. More environments will follow. The sources do not say whether special terms apply in Denmark or the EU.

What's smart about it

Fusion uses three working modes. In Single, one model solves the task. In Cascade, a cheaper model drafts a solution and a stronger model takes over if the quality check requires it. In Critique, one model drafts a solution, another reviews it, and the first one revises it.

The idea is simple: use the expensive models when they are needed, not all the time. Companies also get control. Administrators can restrict which models Fusion is allowed to use and make Fusion the default for their employees.

Cheaper or better?

Users pay for each model Fusion uses at that model's regular rate. Users with paid Copilot plans get a 30 percent discount on model costs for a limited period. The sources do not say when the discount expires, and a monthly price for Fusion has not been disclosed.

Artificial Analysis measured that Fusion on its max setting cost $5.59 per task, compared with $11.42 for Claude Code with Opus 5.5. Fusion used 7.2 million tokens (chunks of text) per task, roughly half as many as Opus 5.5 in Claude Code. That applies to Fusion Max in this particular comparison, not to pricing in general.

In Coding Agent Index v2.0, Fusion scored 71 on Terminal-Bench 4.0, which was first place. On FrontierCode it scored 51 and came in third behind Opus 5.5, which scored 56. On SWE-Atlas-QnA-AA it landed in the middle of the pack with 57. Artificial Analysis concludes that Fusion delivers near-frontier performance at significantly lower cost, meaning almost the same quality as the very best models.

GitHub itself says Fusion Max scored 60 points on the same index at roughly half the average API cost per task, with half as many tokens and 40 percent less time.

Who it's good for

Fusion suits developers and companies that use AI for code every day and want to save money without having to choose a model every time. Standard covers routine work, while High and Max are meant for the hard tasks.

For IT managers, the administrator controls are a plus, since they can limit which models are used. The numbers show that Fusion is strong on terminal tasks, but it doesn't beat Opus 5.5 across the board.