GitHub has made Copilot Fusion generally available, giving developers a single option that decides which artificial intelligence models handle a coding task and how they work together. The company announced the launch at its Universe developer conference on Thursday, October 29, 2026, presenting Fusion as a shift away from asking each developer to pick the right model for every prompt.
Developers who select Fusion in the model picker are choosing a system rather than a specific model. For each prompt, it assesses what the task demands, whether that is reasoning, code generation, debugging or tool use, and routes the work accordingly. Users only set an effort level: Standard for everyday coding, High for tougher problems, and Max for the most demanding tasks.
Key Takeaways
- GitHub has made Copilot Fusion generally available, announced at its Universe developer conference on October 29, 2026.
- Instead of picking a single model, developers pick an effort level — Standard, High or Max — and Fusion routes each prompt to the models it judges suitable.
- GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, at about half the average API cost per task reported for the comparison model.
How Fusion Picks Its Models
Fusion uses three workflows, depending on the task. In the simplest, one model solves the task directly. In a cascade, a cheaper model drafts an answer and a quality check decides whether to accept it or escalate to a stronger model. In critique mode, one model drafts, another reviews, and the first revises its work in light of that review.
The idea is to stop developers from paying for the most powerful model on every job, and from using too weak a model on a job that needs more. GitHub’s own example, reported from the announcement, is a familiar one: a developer picks a powerful model for a hard task and then forgets to switch back for routine work, so ordinary jobs keep running at premium cost. Fusion is designed to make that switching automatic, within the limits set by the team.
That question of cost is now central to how companies judge AI tools. Recent reporting on businesses buying AI faster than they are making it pay shows why a tool that promises the same quality for less money will get close attention — and why its claims will also be tested carefully in practice.
Developers who select Fusion in the model picker are choosing a system rather than a specific model.
What GitHub Claims on Cost and Performance
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. These are GitHub’s own figures, drawn from its announcement, and they describe a benchmark result rather than a guarantee for any particular codebase or team.
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. That wider claim, like the benchmark score, will be judged by developers once they can compare Fusion against the models they already use on their own work. Benchmarks are a useful guide to how AI models are measured, but they do not replace testing on real projects, where code style, existing systems and team habits all affect the result.
On billing, users pay for each model Fusion uses at that model’s standard rate. For a limited period, paid Copilot subscribers get a 30 percent discount on model costs when using Fusion. That discount is presented as a launch offer, not a permanent price, so teams comparing costs should check what they would pay once it ends.
- Fusion offers three effort levels: Standard, High and Max.
- Its three workflows are direct solving, a cheaper-model cascade with escalation, and draft-and-critique revision.
- GitHub reports Fusion Max at 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max.
- Paid Copilot subscribers get a 30 percent discount on model costs with Fusion for a limited period.
- Fusion is available in Visual Studio Code, the GitHub Copilot app, the Copilot command-line tool and JetBrains IDEs.
Controls for Teams and Enterprises
For larger teams, the controls may matter as much as the routing. Administrators decide which models Fusion can draw from, and Fusion only uses models enabled under an organisation’s existing policies. Enterprise administrators can also make Fusion the default for their developers, rather than leaving each person to choose.
AT&T’s Hari Pokhrel, a principal software engineer, said Fusion Standard has cut unnecessary spending across the company, according to the announcement coverage. He noted that developers often pick a powerful model for a hard task and forget to switch back for routine work — the exact pattern Fusion is built to remove.
Fusion combines Project HydraFusion, which began as a research preview, with Copilot’s automatic model selection. Preview users will be moved to the matching Fusion effort levels, so existing preview users should see their settings carry over rather than starting from scratch.
The Bigger Push at Universe
The launch came alongside other announcements at Universe. GitHub Enterprise customers can now use every Copilot feature without a separate subscription, paying only for the AI credits they consume. GitHub also released a Copilot app that runs agent work across multiple repositories, a step towards agents that can work across a whole codebase rather than a single file at a time.
GitHub’s annual Octoverse report, published the same day, showed how fast AI agents are spreading through software work. 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, with India accounting for 32 million of them. That scale is why a change to how models are chosen matters beyond a single tool: even small savings per task add up across millions of developers, and so do small quality differences.
The same pressure to prove real, useful work — rather than simply counting activity — is behind other recent efforts to measure how much real work AI systems actually do. For coding tools, the test will be whether Fusion’s routing holds up on messy, real-world code, not just on benchmark tasks.
What to Watch Next
What to watch next is what the article has already set out: whether developers find Fusion’s effort levels match their real work, whether the cost savings survive once the limited-period 30 percent discount ends, and whether independent testing backs up GitHub’s benchmark and cost claims. The other tests are whether enterprise teams adopt Fusion as the default, whether support reaches more clients beyond Visual Studio Code, the Copilot app, the command-line tool and JetBrains IDEs, and whether the Octoverse trend towards agent-drafted pull requests keeps rising. Until those questions are answered, Fusion stands as a clear bet by GitHub: that the next step for AI coding is not a bigger single model, but a system smart enough to choose the right model for each job.