A developer explores GitHub Copilot Fusion on dual monitors, with orchestration options and a benchmark score displayed.
GitHub has widened Copilot Fusion in three ways. It now shows up on more Copilot surfaces, it has new models behind its effort tiers, and it has an entry on the public Terminal-Bench 4.1 leaderboard. Here is what the changelog says, what the leaderboard shows, and where the marketing math needs a footnote.

A developer explores GitHub Copilot Fusion on dual monitors, with orchestration options and a benchmark score displayed. What actually changed​

Fusion went generally available on October 28. GitHub described it as a system that routes and orchestrates models, not a model in its own right. That framing matters, so I'll set it aside and use GitHub's own description of how it works. You pick Copilot Fusion in the model picker and choose an effort level. Fusion then decides which models to use for each prompt.

The November 17 update does three things:

  • More places to use it. Fusion can now be selected 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 mobile apps is "coming shortly", so it hasn't shipped there yet.
  • New models behind the tiers. Claude Sonnet 5.5 goes to Fusion Max, GPT-6.1 Terra to Fusion High, and MAI-Code-2 to Fusion Standard.
  • A public benchmark entry. Copilot Fusion now appears on the Terminal-Bench 4.1 leaderboards hosted by Stanford, Harbor, and the Laude Institute.

At launch, Fusion was available on all Copilot plans in VS Code, the GitHub Copilot app, Copilot CLI, and JetBrains IDEs. Visual Studio, the web, and mobile are the new additions. The update doesn't give per-client version requirements or plan eligibility beyond that.

How the effort tiers work​

According to GitHub's launch post, the tiers are:

  • Standard prioritizes versatile models for everyday coding.
  • High prioritizes capable models for harder tasks and deeper reasoning.
  • Max prioritizes the most capable models for the most demanding work.

Behind the tiers, Fusion can use one of three workflows. In Single, one model solves the task. In Cascade, an efficient model drafts and a quality check either accepts the result or escalates to a stronger model. In Critique, one model drafts, another reviews, and the first revises. The new model additions don't tell you which workflow runs for a given prompt, and GitHub hasn't said which was used for the benchmark run.

Also note that the new models are additions to the tiers. You still don't choose them directly. The announcement doesn't name the vendor behind GPT-6.1 Terra or MAI-Code-2. It also doesn't say whether the additions are on by default for every account.

The Terminal-Bench 4.1 numbers​

GitHub reports that Fusion Max resolved 71.2% of tasks. It calls that 5.4 percentage points ahead of Opus 5.5 and Fable 5.2, at 88% lower cost than Fable 5.2 and 57% lower than Opus 5.5.

The public leaderboard data backs up the headline score. The Harbor leaderboard submission pull request records the Fusion entry as max effort, run in GitHub Copilot, with 71.2% ± 2.9% accuracy, 3.3B tokens, and a $2.0k cost. The submission bot's trial summary shows 330 trials, 328 of them without error, one verifier timeout, and one agent timeout. The bot also reports that no reward-hacking trials were disqualified.

The leaderboard page shows Opus 5.5 (max) in Claude Code at 65.8% ± 3.1%, with 8.0B tokens and $4.7k. Fable 5.2 (max) in Claude Code is at 65.8% ± 2.8%, with 15.7B tokens and $16.5k. Those totals are roughly consistent with GitHub's 57% and 88% savings claims. They are rounded, though, so I'd quote GitHub's percentages as GitHub's.

Caveats worth keeping in mind​

  • The intervals overlap. The ± figures are 95% confidence intervals. Fusion's range is roughly 68.3% to 74.1%, and Opus and Fable top out near 68.9% and 68.6%. The ranking is real, but the leaderboard alone doesn't prove the gap is statistically significant.
  • It compares systems, not models. The entries are different agent harnesses and configurations: Copilot, Claude Code, and others. This is not a head-to-head of isolated model weights.
  • The cost is a benchmark cost. The $2.0k figure is the total for the benchmark run. It isn't what a subscriber would pay, and it says nothing about the limited-time 30% Fusion discount GitHub announced in October for paid plans.
  • Terminal tasks are one slice of coding work. Independent testing from Artificial Analysis around launch showed a mixed picture. In its Coding Agent Index v2.0, Fusion Max in Copilot CLI scored 60 at $5.59 per task, against 62 at $11.42 for Claude Code with Opus 5.5 Max. On FrontierCode, Claude Code with Opus 5.5 Max and Devin Fusion both scored higher than Fusion. A team doing repository-level changes may therefore see a different trade-off than the terminal headline suggests.

What admins should know​

GitHub's launch post says Fusion only uses models enabled in your enterprise and organization model policies. Enterprise admins can see which models Fusion picks from, set each model's effective policy state, and disable a model for Fusion without disabling it for everyone. Admins can also make Fusion, at any effort level, the default for Copilot users through enterprise managed settings.

That matters now because the new tiers add Claude Sonnet 5.5, GPT-6.1 Terra, and MAI-Code-2. If your compliance rules restrict certain vendors, check your model policies. The November update doesn't introduce a new policy procedure, so I won't invent one.

Billing is also model-based. GitHub says you're billed for each model Fusion uses, at that model's standard rate. The October post says paid-plan users get a 30% discount on model costs for a limited time. Neither announcement gives an end date.

Practical takeaways​

  1. Visual Studio users can now try Fusion by updating to the latest version, opening the model picker, selecting Copilot Fusion, and choosing an effort level.
  2. Mobile users need the beta apps for now. Production apps are still pending.
  3. Cost-conscious teams should run their own representative tasks before trusting any benchmark. The 71.2% and the cost gap come from one terminal-focused suite.
  4. Auditors and reproducibility-minded teams should remember that routing means the model used can vary by prompt. Fusion's choices are not pinned.

Bottom line​

This is a solid expansion for Microsoft's developer ecosystem, bringing Fusion to Visual Studio and the web. The public leaderboard entry is a welcome step toward verifiable claims. Still, a first-place score with overlapping error bars, on a single benchmark family, is a data point and not a verdict.

 

References

  1. GitHub Copilot Fusion updates: more availability, benchmarks, and models GitHub Changelog 2026-11-17T18:04:48+00:00
  2. Leaderboard Submission: Fusion + GitHub Copilot by github-actions(bot) · Pull Request #2848 · harbor-framework/terminal-bench · GitHub github.com
  3. Route and orchestrate models with GitHub Copilot Fusion - GitHub Changelog github.blog