Copilot Fusion tops Terminal-Bench 4.1 at a fraction of the cost
GitHub says Fusion Max resolved 71.2% of Terminal-Bench 4.1 tasks, 5.4 points above Opus 5.5 and Fable 5.2, at 57% to 88% lower cost. Fusion also reaches the web, Visual Studio and mobile.

GitHub, owned by Microsoft, says Copilot Fusion Max resolved 71.2% of tasks on the public Terminal-Bench 4.1 leaderboard, 5.4 percentage points ahead of Opus 5.5 and Fable 5.2. In its changelog post, GitHub also extends Fusion to more surfaces and adds three models.
A leaderboard win that GitHub says is also cheaper
Terminal-Bench 4.1 tests how well AI agents finish hard, realistic tasks in a terminal. The leaderboards are hosted by Stanford, Harbor and the Laude Institute, so the run follows shared evaluation rules instead of a vendor’s own harness. As of the leaderboard page we read, Copilot Fusion (max) leads at 71.2% with a margin of 2.9 points. Opus 5.5 (max) and Fable 5.2 (max), both running in Claude Code, each score 65.8%. Gemini 4 Carbon (high) in Antigravity CLI follows at 64.5%, then Sonnet 5.5 (max) at 63.9%.

GitHub adds a cost claim: Fusion Max cost 88% less than Fable 5.2 and 57% less than Opus 5.5 on the same benchmark. The post gives no dollar figures or token counts behind those percentages, so readers cannot reproduce the comparison from the announcement alone.
Web, Visual Studio and mobile join the list
Fusion can now be chosen in the model picker on the latest github.com, Visual Studio, and the GitHub Mobile beta apps for iOS and Android. GitHub says support in the production mobile apps is coming shortly, without a date. The user sets an effort level and Fusion picks and coordinates the models for each task.

Three models are added. Claude Sonnet 5.5 joins Fusion Max, GPT-6.1 Terra joins Fusion High, and MAI-Code-2 joins Fusion Standard. That means the top tier now blends models from different vendors, including one from Microsoft itself.
Fusion was already generally available in VS Code, Copilot CLI, the GitHub Copilot app, JetBrains IDEs and Copilot in the cloud, according to GitHub’s documentation. Our earlier report on JetBrains covers another surface that received updates this week.

What the benchmark does and does not tell developers
The angle that matters most in practice is who can use this and at what price. GitHub’s docs say Fusion adds no charge of its own: each model call is billed at that model’s normal rate. Standard is open to all plans, including Free, while High and Max need Pro+, Max, Business or Enterprise. Because those upper tiers can involve several models, a single task may use more AI credits than on Standard. The docs also mention a 30% discount on model costs for paid plans through November 30, 2026, which is less than two weeks away.
The docs describe how Fusion works: it routes each prompt to a single model, to a cascade where a cheaper model drafts and a judge accepts or escalates, or to a critique pattern where several models review a draft. This helps explain how a router can cost less than a single frontier model: it only adds calls when a task seems to need them. The docs also warn that if a discarded draft already edited files, those edits are not undone automatically.
Several things remain unestablished. The 71.2% and cost figures are GitHub’s own claims about a system that mixes models, while the competing entries on the board are single models inside their vendors’ agents, so the comparison is between whole systems, not models alone. The leaderboard lists confidence margins of about 3 points, and the second to fifth places sit within a few points of each other. The post does not say how the discount on cost was measured, or whether the new models changed the leaderboard run. Both the Fusion tiers and the model pool may change, per the docs.
For teams already using Copilot, the practical change is a choice rather than a migration: Fusion is one more entry in the picker, now also on the web and in Visual Studio. GitHub shipped related Copilot updates the same day, including Evals for testing skills and MCP tools, which gives teams a way to check agent behavior on their own work rather than relying on a public leaderboard.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Leo MartinsAI Tools, Prompts & Practical AI ExpertLeo Martins is the AI Tools & Practical AI Expert at Aiholics, focused on helping readers use artificial intelligence to improve productivity, creativity and everyday work.
You may also like

Copilot retires Claude Opus 4.8, Opus 5 and GPT-5.6 Terra from its model picker

Copilot app now works with GitHub Enterprise Server, minus cloud sessions

GPT-6.1 Terra and three more Copilot models hit errors, blamed on a provider
