AI Aulis

AI NEWS IN ENGLISH · 18 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

32 stories · next issue tomorrow morning · How the paper is made →

← Issue 18 November 2026Paper

Copilot Fusion Max solved 71.2 per cent of Terminal-Bench 4.1 tasks, GitHub says

AI-generated contentWritten by AI

github.blog, x.com, youtube.com, blogs.windows.com · 7 sources · Sources published: 17 November 2026

GitHub said Copilot Fusion Max, which picks and combines models for each task, solved 5.4 percentage points more Terminal-Bench 4.1 tasks than Opus 5.5 and Fable 5.2 run with Claude Code, and cost less than either.

Copilot Fusion is not a new model but an option in Copilot’s model picker. Users choose a tier rather than a model: Standard, High or Max. Fusion then decides for each prompt which model does the work and at what reasoning effort.

Standard selects one versatile model, according to the subtitles of GitHub’s demo video. In preview, that can also be a local model running on the user’s machine. High and Max can also combine models. In a cascade, a lightweight model drafts a solution, and a judge either accepts it or escalates the work to a stronger model. In a critique, several models review the draft in parallel, and one model fixes it based on their feedback. In the video’s example, GPT-6.1 Sol implemented a feature, GPT-6 Luna reviewed it and Opus 5.5 made the fixes.

Fusion Max solved 71.2 per cent of Terminal-Bench 4.1 tasks, GitHub said in a changelog published on 17 November 2026. The benchmark measures how well AI agents handle hard command-line tasks. Stanford, Harbor and the Laude Institute maintain its public leaderboard. GitHub said Fusion was 5.4 percentage points ahead of Opus 5.5 and Fable 5.2. It also cost 88 per cent less than Fable 5.2 and 57 per cent less than Opus 5.5.

The comparison does not pit models alone against each other. Each row on the leaderboard pairs a model with a harness: the program that drives the agent, gives the model its tools and runs its steps. The image in GitHub’s post on X shows Fusion run with GitHub Copilot, and Opus and Fable with Claude Code. Fusion itself can use Opus 5.5, for example.

Nor was Fusion the cheapest in the image’s table. GPT-6.1 Sol run with Codex and Gemini 4 Carbon run with Antigravity CLI, for example, cost less but solved fewer tasks. The table also gives an uncertainty interval for each result, and Fusion’s interval overlaps slightly with those of Opus 5.5 and Fable 5.2.

This is therefore the result of a single benchmark. Users can measure their own Copilot setup with Copilot Evals, which GitHub released in public preview the same day. The tool lets users build a repeatable test for a skill, their own agent or an MCP tool and compare results before and after a change. GitHub said the tool is needed because AI output varies from run to run, so one successful answer does not yet show that a change works. Evals is available as a VS Code extension, as a Copilot CLI plugin and in the Copilot app for Individual, Business and Enterprise plans.

Users can now select Fusion in the latest versions of github.com, Visual Studio and the GitHub Mobile beta apps. It will reach the production versions of the mobile apps soon. The Max tier gained Claude Sonnet 5.5, High gained GPT-6.1 Terra and Standard gained MAI-Code-2. Microsoft said that, in preview, Fusion can route coding tasks to local models such as MAI Code 1.1 Flash. The feature is available in the Copilot app, the CLI and VS Code on Windows PCs with NVIDIA RTX Spark or AMD Ryzen AI Max processors. GitHub Enterprise Cloud companies using Enterprise Managed Users accounts can also send Copilot agent activity to Microsoft Agent 365 for monitoring.

Previous coverage

Report an error in this story