GitHub Copilot Fusion Max gets 71.2% on Terminal-Bench 4.1
GitHub tells that Fusion Max has 5.4 percentage points more than two other models, at 57% to 88% less cost.
GitHub adds Copilot Fusion to the public Terminal-Bench 4.1 leaderboards of Stanford, Harbor and the Laude Institute. Terminal-Bench measures how many hard tasks that AI agents complete in a terminal. GitHub tells that Fusion Max has a resolution rate of 71.2%. Fusion is an item in the model picker that selects and controls models for each task.
New items
You can select Fusion on the GitHub website, in Visual Studio and in the GitHub Mobile beta apps. The production mobile apps get it in a short time. You set the level and Copilot does the other work. GitHub gives these models for the levels:
- Fusion Max: Claude Sonnet 5.5
- Fusion High: GPT-6.1 Terra
- Fusion Standard: MAI-Code-2
The numbers
GitHub tells that Fusion Max has 5.4 percentage points more than Opus 5.5 and Fable 5.2. It tells that the cost is 88% less than Fable 5.2 and 57% less than Opus 5.5.
Unknown points
GitHub gives the cost numbers and does not tell how it measures them. It gives no results for High and Standard.
Insight: GitHub now claims to lead, not only to be close
In October, GitHub tells that Fusion Max is less than two points from Opus 5.5 Max at half the cost. It tells at this time that Max is above two other models on an external board. Some Fusion models are not from GitHub. GitHub seems to put Fusion as the router for models, and it controls the customer and the price. Monitor if model makers put their agents on the same board.
Sources
Posted