wasl.tech

GitHub Copilot Fusion Max Tops Terminal-Bench 4.1, Beating Claude and Fable for Less

Every developer who has opened a coding assistant lately has met the same small, paralysing moment. A dropdown full of model names. A dozen options, each with its own strengths, prices and quirks.

Pick the wrong one and you burn tokens on a job a cheaper model could have done in seconds. Pick the right one and you still wonder whether another would have done it better.

GitHub Copilot Fusion exists to end that moment. It does not ask you to choose a model at all. It chooses for you, task by task, and on November 17 it got a result worth talking about.

According to GitHub, Copilot Fusion Max has now topped a public leaderboard that the industry actually watches, beating two of the most respected frontier models while costing far less to run.

The scoreboard in question is Terminal-Bench 4.1, a public benchmark hosted by Stanford, Harbor and the Laude Institute. It measures how well AI agents complete hard, realistic tasks inside a terminal.

Copilot Fusion Max reached a 71.2% resolution rate there, 5.4 percentage points ahead of both Claude Opus 5.5 and Claude Fable 5.2, which each scored 65.8%.

The cost side of the story is arguably the bigger one. GitHub says Fusion Max cost 88% less than Fable 5.2 and 57% less than Opus 5.5 on the same tasks.

That is not a rounding error. It is the difference between a tool a team can afford to run all day and one it has to ration.

What GitHub Copilot Fusion Actually Is

The first thing to understand is that Copilot Fusion is not a model. It is a router. When you give it a task, it looks at what the job needs and picks one or more models from a shared pool of 13 to do the work.

Sometimes it hands a job to a single model. Sometimes it splits the work or escalates it to a stronger model when the first attempt falls short.

That is the whole idea. No single model is best at everything, and the best model for a quick fix is rarely the best for a deep debugging session. Fusion matches the tool to the task instead of making the developer do it.

GitHub made it generally available on October 28, 2026, at its Universe conference, replacing an earlier research preview called HydraFusion.

Fusion comes in three effort levels. Standard is the lightest touch, High goes further, and Max is the deepest reasoning tier. The tier you can use depends on your plan.

Fusion Max needs Copilot Pro+ at $39 a month, while the $10 Pro plan and the free tier get Fusion Standard only. Fusion does not bill separately: you pay each model’s standard rate as it is used, and paid plans get a 30% discount on Fusion’s model costs through November 30, 2026.

Copilot Fusion Max Tops Terminal-Bench 4.1

The November 17 update is what put Fusion on the public leaderboard in the first place. GitHub says Copilot Fusion is now listed on the Terminal-Bench 4.1 public leaderboards, and the headline number is the one that will travel furthest.

Fusion Max resolved 71.2% of the benchmark’s tasks, against 65.8% for both Opus 5.5 and Fable 5.2.

The leaderboard itself lists the score as 71.2% with a margin of plus or minus 2.9%, so the lead is real but not enormous. What makes it striking is the combination with cost.

GitHub’s own calculations, based on the costs listed on the leaderboard, put Fusion Max at 88% below Fable 5.2 and 57% below Opus 5.5 per task. In other words, the router did better work for dramatically less money. KeoCodes verified the figures against GitHub’s changelog and the public leaderboard.

This is not Fusion’s first strong showing on an independent benchmark. At its launch in late October, GitHub reported that Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Opus 5.5 Max.

That result came at about half the API cost per task, half the tokens and 40% less time. The new Terminal-Bench 4.1 result is a second, independent data point, and it tells the same story: a system that routes between models can match or beat a single frontier model, because it spends its budget where it matters and saves it where it does not.

New Models Join Each Copilot Fusion Tier

The November 17 update was not only about benchmarks. GitHub also added new models to each Fusion tier, widening the pool the router can draw from.

Claude Sonnet 5.5 joins Fusion Max, giving the deepest tier one of Anthropic’s newest models. GPT-6.1 Terra joins Fusion High, and Microsoft’s own MAI Code 2 joins Fusion Standard, putting a capable homegrown coding model in the cheapest, most widely available tier.

The same day brought a stack of other Copilot changes. GitHub added OpenAI’s brand-new GPT-6.2 Astra and GPT-6.2 Sol models to Copilot for Pro+ subscribers and above, along with MAI Code 2 and a lighter MAI Code 2 Flash.

In GitHub’s own early testing, GPT-6.2 Sol caught nearly twice as many real issues in code review as GPT-6.1 Sol, while Astra caught significantly more than GPT-6 Astra using noticeably fewer tokens.

Copilot Fusion Reaches the Web, Visual Studio and Mobile

Fusion is also showing up in more places. Until now it lived mainly in VS Code, the Copilot command-line tool, the Copilot app and JetBrains IDEs. The November update brings it to github.com itself, to Visual Studio and to the GitHub Mobile beta apps, with the production mobile app coming shortly.

That expansion matters because of where developers actually work. A lot of coding assistance now happens in the browser, in a pull request or on a phone between meetings, not only inside an editor. Putting the router on the web and on mobile means the same model-choosing logic follows the developer everywhere GitHub does.

Why the Copilot Fusion Result Matters for Developers

For most developers, the practical question is not which model tops a leaderboard. It is what this costs and whether it saves time. Fusion’s pitch is that it answers both questions at once.

You stop paying the premium for a frontier model on tasks that do not need one, and you stop getting weak results from a cheap model on tasks that do.

The cost figures deserve a careful read. They are GitHub’s own calculations, drawn from the costs published on the leaderboard, and they describe API cost per task rather than what shows up on a Copilot bill.

Real-world savings will depend on your plan, your usage and how often Fusion escalates a task to a stronger model. Still, the direction is clear, and it matches the earlier Artificial Analysis result.

There is also a competitive angle. Anthropic’s Claude models have set the pace for coding assistants for much of the year, and OpenAI’s GPT-6.2 launch this week was aimed squarely at professional work. Google’s Gemini 4 Carbon is the other model in that top tier.

GitHub’s answer is not a bigger model of its own. It is a system that uses everybody’s models, including its rivals’, and charges less for the privilege. If Fusion’s benchmark numbers hold up in everyday use, GitHub has made a strong case that orchestration, not raw model size, is where the advantage now lives.

What to Watch Next for Copilot Fusion

Two dates matter in the short term. The 30% discount on Fusion’s model costs for paid plans runs through November 30, 2026, so the cheapest time to try the Max tier is right now. The production GitHub Mobile app, with Fusion built in, is also due shortly after the beta.

The longer question is whether the leaderboard lead survives contact with real codebases. Benchmarks are clean and repeatable. Real projects are messy, full of private context, legacy code and half-written requirements.

For now, though, the headline stands: a router that picks its models beat two of the best models in the world on a public test, at a fraction of the price. That is the kind of result that changes how teams think about paying for AI coding tools.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top