GitHub Universe: Copilot stops selling you a model and starts picking one for you
GitHub used Universe on October 28 to change what you are buying. You no longer pick a model. Copilot Fusion looks at each task and "chooses how to execute across one or more models," drawing on Opus 5.5, Sonnet 5.5, GPT-6 Astra and GPT-6 Sol. You pick an effort level, and GitHub picks the engine.
The number GitHub led with: on the Artificial Analysis Coding Agent Index, Fusion Max lands within two points of Opus 5.5 Max "at half the cost and 40% less time per task." Artificial Analysis posted its own figures the same hour: $5.59 per task against $11.42 for Claude Code running Opus 5.5 at max effort, about 7.2 million tokens per task, roughly half of Claude Code's, and first place on Terminal-Bench 4.0 at 71. A router built on top of other labs' models now beats every single-model harness on one of the hardest agent benchmarks. That says as much about the harness as about any model underneath.
The rest of the launch is about running many agents at once. The new Copilot app is "designed from the ground up for agentic software delivery." Copilot projects run parallel managed sessions across several repos, and a session carries over between the app, VS Code, the CLI and the web. Agentic Workflows is now generally available. Copilot Automations "can learn from their outcomes and propose improvements for you to review and approve." And the pricing move may matter most: GitHub Enterprise customers now get every Copilot feature without a separate seat license and "pay for the AI credits you use."
The Octoverse numbers released alongside show why GitHub is rebuilding its plumbing. Commits hit 7.4 billion a month, up from 2.9 billion in August. Repos with an agent configuration file log eight times as many commits as new repos without one. GitHub says it is moving to a new Git architecture for "agent-scale" throughput.
Our read: this is the sixth serious entrant in model routing, and the biggest by distribution. When the platform that hosts your code also decides which model writes it, the labs are selling to GitHub, not to you. The catch is that the cost and quality numbers so far are GitHub's and one independent index. Teams should check them against their own repos before they let a router choose.
Universe post: https://github.blog/news-insights/company-news/github-universe-2026-evolving-every-layer-of-the-stack/ Octoverse: https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/
← Back to all articles
The number GitHub led with: on the Artificial Analysis Coding Agent Index, Fusion Max lands within two points of Opus 5.5 Max "at half the cost and 40% less time per task." Artificial Analysis posted its own figures the same hour: $5.59 per task against $11.42 for Claude Code running Opus 5.5 at max effort, about 7.2 million tokens per task, roughly half of Claude Code's, and first place on Terminal-Bench 4.0 at 71. A router built on top of other labs' models now beats every single-model harness on one of the hardest agent benchmarks. That says as much about the harness as about any model underneath.
The rest of the launch is about running many agents at once. The new Copilot app is "designed from the ground up for agentic software delivery." Copilot projects run parallel managed sessions across several repos, and a session carries over between the app, VS Code, the CLI and the web. Agentic Workflows is now generally available. Copilot Automations "can learn from their outcomes and propose improvements for you to review and approve." And the pricing move may matter most: GitHub Enterprise customers now get every Copilot feature without a separate seat license and "pay for the AI credits you use."
The Octoverse numbers released alongside show why GitHub is rebuilding its plumbing. Commits hit 7.4 billion a month, up from 2.9 billion in August. Repos with an agent configuration file log eight times as many commits as new repos without one. GitHub says it is moving to a new Git architecture for "agent-scale" throughput.
Our read: this is the sixth serious entrant in model routing, and the biggest by distribution. When the platform that hosts your code also decides which model writes it, the labs are selling to GitHub, not to you. The catch is that the cost and quality numbers so far are GitHub's and one independent index. Teams should check them against their own repos before they let a router choose.
Universe post: https://github.blog/news-insights/company-news/github-universe-2026-evolving-every-layer-of-the-stack/ Octoverse: https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/
Comments