Unlike a standalone AI model, Copilot Fusion acts as an orchestration system that can use one model or coordinate several during a single task. Developers select Fusion through the Copilot model picker and choose among three effort levels: Standard for everyday coding, High for more challenging assignments and Max for tasks requiring the most advanced reasoning.
Fusion then determines how to distribute the work using three execution methods.
The simplest approach sends a request directly to a selected model. For more demanding assignments, Fusion can begin with a less expensive model and evaluate whether its response is sufficient before escalating to a more capable alternative. A third method uses one model to produce an initial result, another to review it and the original model to make revisions.
GitHub Chief Product Officer Mario Rodriguez told VentureBeat that the goal is to reduce the burden of choosing models while improving the economics of AI-assisted development.
“It's not a model. It's an orchestration engine at runtime,” Rodriguez said.
The system can draw on models from multiple providers, including Anthropic and OpenAI, subject to the models an organization permits. Enterprise administrators can control which models Fusion uses and establish it as the default option for developers.
Early benchmark results suggest that the approach can deliver substantial cost savings without a comparable reduction in coding performance.
In testing conducted by Artificial Analysis, Copilot Fusion Max scored 60 on the Coding Agent Index, compared with approximately 62 for Claude Code running Anthropic's Opus 5.5 Max configuration.
Fusion's average cost was $5.59 per task, versus $11.42 for the Claude Code configuration, representing a reduction of approximately 51%. It also consumed roughly half as many tokens.
GitHub separately reported that Fusion completed the benchmark assignments about 40% faster. Microsoft CEO Satya Nadella cited potential cost reductions of up to 77%, although that figure came from a different internal evaluation rather than the Artificial Analysis comparison.
The performance differences also varied by task. Fusion achieved a score of 71 on Terminal-Bench 4.0, but trailed Claude Code and Devin Fusion on FrontierCode, an evaluation focused on modifying software repositories.
Those results indicate that Fusion's overall performance is competitive, although it does not outperform alternative coding systems across every benchmark. The reported savings also reflect controlled evaluations rather than the total cost of deploying AI coding tools within an organization.
GitHub charges for the individual models Fusion uses at their standard rates. For a limited period, paid Copilot subscribers receive a 30% discount on model costs when using Fusion.
The system is available across all Copilot plans in Visual Studio Code, JetBrains IDEs, GitHub Copilot CLI and the GitHub Copilot app, with additional client support planned.
Beyond Fusion, GitHub announced changes to how enterprise customers purchase and manage Copilot.
GitHub Enterprise customers can now access the full range of Copilot features without purchasing a separate subscription, instead paying according to their AI credit consumption. The arrangement is initially available through GitHub's sales organization, with self-service purchasing expected later.
Existing Copilot Business and Copilot Enterprise customers can retain their current subscriptions or move to the new arrangement.
GitHub also introduced additional capabilities for coordinating AI agents across development environments.
The GitHub Copilot app brings together agents and tools used across projects, while Copilot projects allow developers to manage parallel agent sessions spanning multiple repositories.
Session continuity enables developers to move an active cloud-based agent session between the Copilot app, Visual Studio Code, the command line and the web without restarting the work.
GitHub has also expanded Copilot automations, allowing them to evaluate previous results and recommend changes for developer approval.
For code review, the company introduced Copilot code review Max, which GitHub says achieved the highest F1 score for identifying high-severity issues on ReviewBench.
These additions reflect GitHub's broader effort to support AI agents throughout the software development process, rather than limiting their role to generating code inside an editor.
The company is also expanding the infrastructure supporting those workloads. According to Microsoft's executive vice president of CoreAI, Jay Parikh, GitHub has added four million CPU cores, 160 petabytes of high-speed storage and 50 terabits per second of network capacity this year.
GitHub is transitioning to a multi-region deployment on Microsoft Azure, with completion expected early next year. The company says the changes will improve platform resilience as automated development activity increases.
That growth is reflected in GitHub's annual Octoverse report, released alongside the announcements.
Developers recorded 17.9 billion code pushes during the 12 months ending September 30, exceeding the combined total from 2024 and 2025. The platform also registered 34.3 billion commits and 1.3 billion merged pull requests.
AI agents are accounting for an increasing share of that activity. In September, more than 40% of pull requests began with an agent-generated draft, compared with 17% in March. GitHub recorded approximately 110 million such pull requests in September alone.
The company cautioned that its measurements likely underestimate agent activity because not every automated contribution can be identified.
GitHub now reports more than 225 million developers on its platform, including 38.3 million in the United States and 32 million in India.
Its data also indicates that developers are increasingly using multiple AI models. Among Copilot users who manually selected models in September, 61% used more than one.
TypeScript remained the platform's most widely used programming language, while Rust recorded the fastest growth among leading languages, doubling over the previous year.
The figures provide context for GitHub's decision to emphasize model orchestration. As developers experiment with different AI systems, Fusion aims to automate those choices while balancing capability, speed and cost.
For enterprise customers, the appeal extends beyond individual coding tasks. GitHub is combining model selection, agent coordination, code review and usage-based purchasing into a broader development platform.
Whether those capabilities produce equivalent savings in production environments will depend on the complexity of projects, the models organizations permit and the additional review and integration work required. The initial benchmarks nevertheless provide evidence that coordinating multiple models can approach the performance of a leading coding agent while substantially reducing the cost of individual tasks.
This analysis is based on reporting from Github & Venture Beat.
Image courtesy of Github.
This article was generated with AI assistance and reviewed for accuracy and quality.
