GitHub Copilot Fusion Arrives as Agents Start 40% of Pull Requests

Agents now start more than 40% of pull requests on GitHub, up from 17% in March. That figure, from GitHub's Octoverse 2026 report, landed the same day the company unveiled GitHub Copilot Fusion at its Universe conference in San Francisco, a system that decides which AI models should handle each part of a coding task.

The two announcements tell one story. GitHub is betting that developers will stop choosing a single model and start handing whole jobs to agents that choose for them. These are the figures behind that bet.

Octoverse 2026 in eight figures

• 40%+: share of pull requests developers started with an agent in September, according to GitHub.
• 17%: the same share in March 2026.
• 225 million+: developers on GitHub worldwide.
• 38.3 million: developers in the US, still the largest community, ahead of India at 32.0 million.
• 17.9 billion: pushes in the 12 months to September 30, more than 2024 and 2025 combined.
• 34.3 billion: commits over the same period.
• 1.3 billion: merged pull requests.
• 61%: share of Copilot users who used more than one model in September.

Agent pull requests more than doubled in six months

Going from 17% to more than 40% in half a year is the sharpest change in the report. The Next Web, which covered the release, puts the September count at about 110 million pull requests started by agents. GitHub has not said how many of those reached production without a person rewriting them, so the share of agent work that actually ships remains an open question.

The volume numbers point the same way. Pushes hit 17.9 billion, more than the previous two years put together, while commits reached 34.3 billion. Part of that growth is human, because the developer base keeps widening, with China at 12.4 million and Brazil at 10.2 million behind the US and India. Still, a jump of this size in agent started work suggests a growing slice of code is now written by software acting for someone.

That shift brings new risks. Researchers have already shown how a planted instruction file can steer coding agents, as we covered in our report on a malicious AGENTS.md file that fooled Codex. More agent pull requests means more code that reviewers need to read carefully.

Why 61% of Copilot users switch models, and what Fusion does about it

The model figure explains why Fusion exists. If most Copilot users already bounce between models, GitHub wants to make that choice automatic. According to GitHub's Universe 2026 announcement, Fusion "evaluates what a task needs across reasoning, code generation, debugging, and tool use," then runs the work across one or more models.

GitHub's headline claim is about cost. The company says Fusion Max scored within two points of Anthropic's Opus 5.5 Max on the Artificial Analysis Coding Agent Index, at half the cost and with 40% less time per task. That is GitHub's own reading of a third party benchmark, and independent testing has not caught up yet.

Routing matters because AI compute is expensive and the money flowing through it keeps growing, as Microsoft's $10 billion OpenAI quarter showed. A tool that sends simple steps to cheaper models could trim those bills for large teams.

Copilot pricing moves to AI credits

The other big change is how companies pay. Starting with Universe, GitHub Enterprise customers can use every Copilot feature without buying a separate subscription. Instead, they pay for the AI credits they consume. For now this runs through GitHub's sales team, with self serve sign up promised later. Existing Copilot Business and Enterprise customers can keep their current plans or switch.

GitHub did not publish dollar prices for credits in its announcement, so teams will have to model costs with their account managers.

Other launches at the event:

• Copilot projects, now native in the GitHub Copilot app.
• Session continuity, so a task started in the app can continue in VS Code, the CLI or the web.
• Copilot automations, which propose improvements for a developer to review and approve.
• Copilot code review Max, which GitHub says posted the highest F1 score on ReviewBench for high severity issues.

4 million new CPU cores behind the agent push

All this agent traffic needs hardware. GitHub says it added 4 million CPU cores, 160 PB of high speed storage and 50 Tb/s of network capacity, and is moving to a multi region Azure setup expected to finish early next year.

Jay Parikh, who wrote GitHub's Universe announcement, framed the stakes there: "I think all the software we've written to date will amount to less than 1% of the software we'll write in the next five years."

If he is even half right, the useful number to watch next year is not how many pull requests agents open, but how many survive review. Tools like Anthropic's proactive Claude Duet agent and recent audits of agent monitoring suggest that question is only getting started. More AI coverage is on our AI news page.