Skip to content

GitHub Copilot Fusion adds models and Terminal-Bench results

GitHub put Copilot Fusion on a public agent benchmark, shipped it to more surfaces, and added three models across its tiers.

Draftship Editorial3 min read

GitHub widened Copilot Fusion on November 17, 2026, adding three models, bringing the feature to more Copilot surfaces, and posting results on a public agent benchmark. The details are in the GitHub Changelog post. Copilot Fusion now appears on Terminal-Bench 4.1 leaderboards hosted by Stanford, Harbor and the Laude Institute, and GitHub reports that Copilot Fusion Max resolved 71.2% of tasks on it.

What the benchmark number covers

Terminal-Bench 4.1, as GitHub describes it, measures how well AI agents complete hard, realistic tasks in a terminal. That scope matters for anyone reading the headline figure: it is a test of agentic work in a shell, not of code completion inside an editor or of chat answers.

GitHub says Fusion Max's 71.2% resolution rate put it 5.4 percentage points ahead of Opus 5.5 and Fable 5.2, which the post calls leading frontier models. It also states that Fusion Max cost 88% less than Fable 5.2 and 57% less than Opus 5.5. Those are relative comparisons only. The post gives no absolute prices, no task counts, and no run configuration. GitHub's stated reason for wanting independent leaderboards is that they let you compare agents on the same tasks under the same evaluation rules, which is the useful part if you are choosing between agents rather than trusting a vendor's own harness.

Where Fusion can be selected

With this release, Copilot Fusion is selectable on the latest versions of github.com, Visual Studio, and the GitHub Mobile beta apps for iOS and Android. GitHub says support in the production GitHub Mobile apps is coming shortly.

The interaction model GitHub describes is a model picker plus a routing layer. You choose Copilot Fusion in the picker, set an effort level, and Fusion then chooses and coordinates models for each task. The post does not break down what the effort levels are or how they change routing.

The three models, by tier

Each addition lands in a different Fusion tier:

  • Claude Sonnet 5.5 goes to Copilot Fusion Max
  • GPT-6.1 Terra goes to Copilot Fusion High
  • MAI-Code-2 goes to Copilot Fusion Standard

Read literally, that suggests the tiers differ in both model roster and intended workload, with Max carrying the heaviest models and Standard the lightest. The post does not say how the tiers map to Copilot plans or seats, so the practical cost of reaching Max is not stated.

What it changes for people building apps

The relevant shift is not a single benchmark number. It is that a vendor is now selling model routing as the product rather than a fixed model choice: you pick a tier and an effort level, and the system decides which model handles which step. For someone shipping an app, that means less direct control over which model touches their code, in exchange for not having to pick per task.

GitHub offers no worked example of a terminal task in the post, so there is no concrete before-and-after to point at here.

Limits and open questions

The post's introduction refers to results on two public Terminal-Bench leaderboards, but only Terminal-Bench 4.1 is named, with three hosting organisations. Whether a second leaderboard is covered is not explained. Also unstated: absolute cost per task, what effort levels exist, how tier access relates to Copilot pricing, when production GitHub Mobile support lands beyond "shortly," and how Fusion picks models inside a tier.

For builders working with AI

Routing and effort controls are becoming standard furniture in AI development tools, and the choice you actually make is which system gets to decide. If you are building an app and want the model choice to stay visible, Draftship's paid plans include a choice of AI model, and using the AI chat and choosing a model covers how that picker works. Draftship is not affiliated with GitHub and does not claim support for Copilot Fusion.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.

AI coding agents

opencode v2.3.2 fixes GPT-6.2 prompt caching billing

opencode v2.3.2 sends the system prompt as a leading developer message on every Responses route so GPT-6.2 caches it, cutting uncached input per step from about 14,000 tokens to 3 on one Copilot model.