GitHub's new AI coding tool nearly matches the best for half the cost

GitHub has launched Copilot Fusion, an AI coding tool that decides which AI models should work on a task. In independent testing, it came within two points of a leading rival while costing about half as much per task. GitHub, a Microsoft subsidiary, unveiled it on October 28 at its Universe conference.

What changed

Fusion is not a single AI model. GitHub's product chief, Mario Rodriguez, called it an orchestration engine, meaning a system that manages several models at once. It looks at the work, then sends it to one model or coordinates a few. Users switch it on and pick one of three effort levels: standard, high or max.

How it works

GitHub's launch documentation describes three patterns for handling a task:

• Single: the work goes straight to one model.
• Cascade: a cheaper model goes first, and a stronger one steps in if needed.
• Critique: one model drafts, another reviews, and the first revises.

The models come from both Anthropic and OpenAI. Which ones are used can vary by request and by the rules an organization sets.

What the tests found

Testing firm Artificial Analysis ran Fusion at its max setting. The average cost was $5.59 per task, compared with $11.42 for Claude Code using Anthropic's Claude Opus 5.5 at its max setting. That is about 51% less. Fusion scored 60 on the firm's coding index, against roughly 62 for Claude Code.

Fusion also used about 7.2 million tokens per task. Tokens are the chunks of text an AI reads and writes, and usage is often billed by them. That is about half the Claude Code setup. Devin Fusion, a competing system, scored 62 at $8.00 per task.

Where it fell short

The results were uneven. Fusion led on Terminal-Bench 4.0 with a 71. On FrontierCode, which focuses on changes to existing code, it scored 51. Claude Code with Opus 5.5 scored 56 and Devin Fusion 54. On a repository question-answering test, Fusion scored 57. A close overall score does not mean it wins every kind of coding job.

Claims to read carefully

GitHub says Fusion Max finished the tasks in about 40% less time. That figure comes from GitHub, and the Artificial Analysis report does not include the time data behind it.

Microsoft CEO Satya Nadella has cited savings of up to 77%. That number comes from a separate GitHub test, where Fusion beat comparable setups by up to 8.7 points on task completion. It is a maximum under different conditions, not the 51% average. The test costs also are not a forecast of what any company will pay under GitHub's new AI-credit purchasing plan.