GitHub Copilot Fusion Picks Your AI Models, Nearing Top Coding Agents at Half the Cost

GitHub has made Copilot Fusion generally available, a system that picks which AI models handle each coding task instead of asking developers to choose one. GitHub says the top setting, Fusion Max, scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Opus 5.5 Max, at about half the average cost per task.

The launch came on Wednesday, October 28, at GitHub Universe in San Francisco. It matters because model choice has become one of the biggest cost and quality decisions in AI-assisted coding, and GitHub is betting most developers would rather hand that decision over.

Key facts at a glance

• What it is: a model router and orchestrator, not a new model. You pick an effort level, and Copilot picks the models.
• Effort levels: Fusion Standard, Fusion High and Fusion Max.
• Benchmark claim: Fusion Max averaged $5.59 per task in testing by Artificial Analysis, against $11.42 for Claude Code with Opus 5.5 Max, roughly 51% less. GitHub says it also used about half as many tokens and took about 40% less time per task.
• Where it works: all Copilot plans in Visual Studio Code, the GitHub Copilot app, Copilot CLI and JetBrains IDEs.
• Billing: each underlying model is charged at its standard rate, with a temporary 30% discount on model costs for paid Copilot plans.

Background: why GitHub built a router

Coding assistants now offer a long menu of models from several AI labs. Picking the right one for each step is tedious, and the wrong pick either wastes money or produces weaker code.

GitHub had been testing two answers to this problem. Project HydraFusion experimented with orchestration, meaning several models working together on one task. Separately, Copilot's auto model selection used a fine-tuned system to pick a single model. Copilot Fusion merges the two.

User behaviour shows why. GitHub says that in September, 61% of Copilot users who chose their own model used more than one. Fusion aims to make that switching automatic.

How Copilot Fusion works

You choose effort, not a model. In the model picker, developers select Copilot Fusion and then an effort level. GitHub says Standard prioritises versatile models for everyday coding, High prioritises more capable models for harder tasks and deeper reasoning, and Max prioritises the most capable models for the most demanding work.

Three workflows run behind the scenes. For each prompt, Fusion decides how models should cooperate:

1. Single: one model solves the task directly.
2. Cascade: an efficient model drafts a response, a quality check decides whether to accept it, and the task escalates to a stronger model if needed.
3. Critique: one model drafts, a second reviews, and the first revises.

Admins keep control. Enterprise administrators can restrict which models Fusion is allowed to use, which matters for companies with data-handling or vendor rules.

Preview users move over automatically. Anyone who had enabled auto, HydraFusion or HydraFusion Max in the research preview is switched to the matching Fusion effort level.

The cost and performance claims

The headline numbers come from two places. GitHub reports the index score, token use and time savings. Artificial Analysis, an independent benchmarking firm, published per-task costs on October 28 showing $5.59 for Fusion Max against $11.42 for the Claude Code configuration.

Billing follows the models used. Fusion does not have its own flat price. Each model it calls is billed at that model's standard rate, so a task routed to cheaper models costs less. The 30% discount for paid plans is a limited-time promotion and is separate from the benchmark savings.

Microsoft's figure is higher still. Microsoft chief executive Satya Nadella said Fusion delivers frontier-level quality 40% faster and at up to 77% lower cost. That "up to" figure is a best case and differs from the roughly 51% saving in the benchmark.

Local models and the wider Copilot harness

Fusion can now route to your own device. In a separate experimental preview, Fusion can choose between supported on-device models and cloud models in Copilot CLI, the Copilot app and VS Code. On supported Windows devices, users can download and configure MAI Code 1.1 Flash through the Windows ML provider. In VS Code, the feature is switched on with the chat.copilot.localModel.enabled setting, and Fusion Standard then chooses between local and cloud models.

The engine underneath has a new name. Microsoft renamed the GitHub Copilot harness, the agent runtime behind Copilot, to the Microsoft Copilot harness. It now powers Copilot experiences across Microsoft and is available through Microsoft Foundry, the company's managed service for building agents. Microsoft says the harness supports more than 20 frontier models plus bring-your-own-key, and that its runtime was recently ported to the Rust language for speed.

The SDK is moving too. The Copilot SDK, available for Python, TypeScript, Go, .NET, Java and Rust, is moving to Microsoft's organisation. GitHub says no immediate changes are required for existing users. Microsoft Copilot Studio makers will see the new name over the coming weeks, with no change to capabilities or billing.

Another perspective

The performance claims deserve some care. The index score, token savings and time savings are GitHub's own figures, and only the per-task cost comparison comes from outside testing. Benchmarks measure a fixed set of tasks, and real projects vary widely in size and difficulty.

The pricing model also shifts some predictability away from developers. Because Fusion bills each model it uses at standard rates, the cost of a task depends on routing decisions the user does not make directly. Teams with strict budgets will want to watch usage reports, especially after the 30% promotional discount ends.

What it means for you

For individual developers, Fusion removes a daily decision. If you have been switching models by hand, try Fusion Standard for routine work and Max for difficult tasks, then compare results and cost against your usual choice.

For team leads and admins, the ability to restrict Fusion's model list is the key control. Set it before rolling Fusion out, so routing stays within the AI providers your company has approved.

For enterprises, GitHub also announced that GitHub Enterprise customers can now use every Copilot feature without buying a separate Copilot subscription, paying instead for the AI credits they use. That offer runs through GitHub's sales team for now, with self-serve promised later. Existing Copilot Business and Enterprise customers can keep their plans.

What happens next

GitHub says support for more clients beyond VS Code, the Copilot app, the CLI and JetBrains is coming soon. Local model routing remains an experimental preview. The 30% Fusion discount is time-limited, though GitHub has not given an end date. Copilot Studio users will see the harness renamed over the coming weeks.

Conclusion

Copilot Fusion turns model choice into a setting. Practical takeaways:

• Update your Copilot client and pick Fusion with an effort level to try it.
• Benchmark it on your own repositories, since the published numbers cover a fixed task set.
• Admins should set model restrictions first, then monitor per-model billing.
• Watch the discount period, because costs follow standard model rates once it ends.