GitHub Copilot Fusion Costs Half as Much as Claude Code, Trailing by 2 Points

GitHub has added a Fusion mode to Copilot. The system decides at runtime which model handles a task: one from Anthropic's lineup or one from OpenAI's. Artificial Analysis, an independent lab that compares AI models and agents under uniform rules, tested Fusion in its Coding Agent Index v2.0. The index averages three benchmarks with equal weight and tests not bare models but ready-made agent-plus-model pairings. The result: Fusion scored 2 points lower than Claude Code with Claude Opus 5.5, but cost half as much. One task came to $5.59, versus $11.42 for Claude Code with Opus 5.5 in max mode.

How Fusion decides who writes the code

GitHub describes Fusion as a real-time orchestration system. It selects and coordinates models for each specific task, balancing quality, price and latency. A complex change in a large repository can go to a top-tier model, while a routine small fix goes to a cheaper one. The user does not choose the model manually.

The pool has four models:

• Claude Opus 5.5, Anthropic's flagship model;
• Claude Sonnet 5.5, a lighter and cheaper Anthropic model;
• GPT-6 Astra, an OpenAI model;
• GPT-6 Sol, another OpenAI model.

Charges follow the rates of the models that actually handled the request. Artificial Analysis data shows no separate markup for orchestration. The test ran in max mode through the GitHub Copilot CLI.

Where Fusion leads and where it trails

The index combines three measurements. Here is how Fusion performed in each:

• Terminal-Bench 4.0 covers tasks the agent solves on the command line. Fusion scored 71 points and took first place, slightly ahead of the rest.
• FrontierCode covers hard programming tasks. Fusion scored 51 and placed third. Ahead of it are Claude Code with Opus 5.5 (max) at 56 points and Devin Fusion, pairing Opus 5.5 (max) with GPT-6 Luna (high), at 54.
• SWE-Atlas-QnA-AA is a block of questions and answers about code. Fusion scored 57 points here, landing between GPT-6 Astra in Codex and Gemini 4 Argon in Antigravity CLI.

A simple average of the three figures comes to about 60. That is consistent with the 2-point gap to the leader. Fusion falls furthest behind on the hardest FrontierCode tasks, where the gap to Claude Code reaches 5 points. On the command line, it outperforms Claude Code.

Why it costs half as much: tokens and price

The price of a task is the volume of tokens (the chunks of text a model reads and writes) multiplied by the rate for each. Fusion is lower on both counts. It uses 7.2M tokens per task. That is roughly half of Claude Code with Opus 5.5 and 30% less than Devin Fusion, which pairs Opus 5.5 (max) with GPT-6 Luna (high). Part of the work goes to inexpensive models, and the costly Opus is called in less often.

Price comparison:

• Copilot Fusion in GitHub Copilot: $5.59 per task;
• Devin Fusion (Opus 5.5 max + GPT-6 Luna high): $8.00 and 62 points in the index;
• Claude Code with Opus 5.5 (max): $11.42.

Fusion is roughly 30% cheaper than Devin Fusion and half the price of Claude Code. Giving up 2 points against the leader and 5 points in FrontierCode is a reasonable price to pay.

What this changes for teams

For most development teams, this is a strong argument for automatic model selection. If an agent solves hundreds of tasks a day, the $5.83 difference per task quickly becomes a serious budget line. Manually picking a model for each type of work is no longer necessary.

There are caveats. For extremely hard tasks of the FrontierCode kind, running Claude Code with Opus 5.5 directly still gives the best result. The data comes from a single index and a single set of tasks, and your own codebase may behave differently. Before switching a team over, run Fusion on your own typical tasks and compare the cost of a solved task, not the price per token.

It is also telling that Fusion mixes models from competing labs. Competition is shifting from individual models to the layer above them: systems that distribute work among those models.