GitHub has a new answer to a question every developer now faces: with so many AI models to choose from, which one should write this piece of code?
At its Universe conference, GitHub announced Copilot Fusion is now generally available. Instead of picking a specific model, developers pick an effort level — and Fusion handles the model choice itself.
What Copilot Fusion actually does
When you select Fusion, you are choosing a system, not a model. For each prompt, Fusion assesses what the task demands — reasoning, code generation, debugging or tool use — and routes the work accordingly.
You set one of three effort levels. Standard is for everyday coding and prioritises efficiency and cost. High brings in more capable models and deeper reasoning for complex tasks. Max is for the most demanding work, where quality matters more than speed or price.
From there, Fusion can work in three ways. In the simplest pattern, a single model completes the task directly. In a cascade, a cheaper model drafts an answer, a quality check judges it, and the work is escalated to a stronger model only if the draft is not good enough. In critique mode, one model drafts, another reviews the draft, and the first model revises its work.
The idea is to stop wasting a powerful, expensive model on a trivial fix — and to stop underpowering a hard problem with a cheap one.
What GitHub claims about cost and quality
GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Anthropic's Claude Code running Opus 5.5 Max. According to the company, it did so at about half the average API cost per task, using about half the tokens and taking about 40 percent less time.
Microsoft chief executive Satya Nadella went further in a LinkedIn post, saying Fusion delivers frontier-level quality 40 percent faster and at up to 77 percent lower cost.
Those are GitHub's own figures, drawn from its announcement, and they have not been independently verified. Treat them as a company claim, not a settled benchmark. Still, the direction is clear: GitHub believes the future of AI coding is not a single best model, but a smart router that spends intelligence where it matters.
On billing, users pay for each model Fusion uses at that model's standard rate, with a 30 percent discount for paid subscribers for a limited period.
Built for teams, not just individual developers
Fusion is available on all Copilot plans in Visual Studio Code, the GitHub Copilot app, the Copilot command-line tool and JetBrains IDEs, with support for more clients coming.
For companies, the controls matter as much as the cost savings. Administrators decide which models Fusion can draw from, and Fusion only uses models enabled under an organisation's existing policies.
Fusion combines Project HydraFusion, which began as a research preview in September, with Copilot's automatic model selection.
This fits a wider pattern in AI tools for work. Just as Oracle AI Workspace brings ChatGPT into a company's apps and data with approvals intact, and Fragment lets AI agents handle enterprise work behind the scenes, Fusion is designed to slot into existing workflows rather than replace them.
Why this launch matters beyond GitHub
The launch came alongside other Universe announcements, including a Copilot app that runs agent work across multiple repositories.
But the most telling number came from GitHub's annual Octoverse report, published the same day. More than 40 percent of pull requests in September began as an agent's draft, up from 17 percent in March. The platform now has more than 225 million developers.
That is why a model router matters. If agents are starting nearly half of all pull requests, the cost of running those agents becomes a central engineering decision. Tools that help developers build with AI, from Bolt to Replit and Lovable, are all competing on the same question: how much intelligence should each task get?
What to watch next
Fusion will be judged on two things. First, whether developers actually trust it to route their work, or whether they keep overriding it to pick a favourite model. Second, whether the cost savings hold up in real codebases, outside GitHub's own benchmarks.
If it works, the model picker could quietly disappear as a daily decision — the way developers stopped choosing a compiler optimisation flag for every build. If it does not, Fusion will be remembered as a clever routing layer that developers bypassed.
Either way, the launch marks a turning point: AI coding tools are no longer just selling a model. They are selling judgement about when to use which model.
Would you let Copilot Fusion choose the model for your next coding task, or would you rather keep that choice in your own hands? Share your view in the comments below.
Comments
Post a Comment