At Half the Price, Copilot Fusion Comes Close to the Top Tier; GitHub Bets on the Entire Pipeline

On day one of GitHub Universe 2026, GitHub focused its keynote on a set of agent workflows from issue to merge. GitHub's press release stated the goal plainly: deliver completed software work—features, bug fixes, and security fixes all count; lines of code do not. Jay Parikh's judgment in the article was even more dramatic: the software written to date may be less than 1% of what will be written in the next five years.

This launch covers multiple stages. The Copilot app connects agents and tools in a project, reducing manual handoffs for users. Copilot Fusion handles multi-model orchestration, selecting model combinations according to task needs. GitHub's numbers: on the Artificial Analysis Coding Agent Index, Fusion Max is 2 points behind Opus 5.5 Max, costs about half, and takes 40% less time per task. These numbers are the easiest to share and the most in need of verification. Dirty repositories, long dependencies, and strange build scripts outside the test set are where the cost equation actually plays out. One commenter said in a reply that if routing also holds up in real repositories, many teams are currently overpaying for agent runs. Another added that paid plans have a discount through the end of November, subject to the official page.

Copilot projects let agents run parallel sessions across multiple repos, with users watching progress and making decisions in one place. Session continuity lets the same cloud agent session continue across the GitHub app, VS Code, CLI, and the web. Copilot automations run AI workflows in repos, triggered manually, by events, or on a schedule, and can learn from results and propose improvements. Code review also adds a Max tier, which GitHub says has the highest F1 score on high-severity issues in ReviewBench. Use Lite and Balanced for everyday review, and Max when deeper investigation is needed. Enterprise customers now do not need to buy a separate Copilot subscription; all Copilot features are available and billed by AI credits.

Some commenters focused on review accountability: when agents start opening PRs across multiple repos, someone still has to be responsible for what changed and why. This reminder is closer to real-world adoption than model leaderboards. One commenter just wanted infrastructure, while another said they wanted to see GitHub stay online. This request is not flashy, but it is very practical. Another commenter asked how Fusion actually splits tasks; that question can only be answered by performance in real projects.

The infrastructure section gave numbers: 4 million new CPU cores, 160 PB of high-speed storage, and 50 Tb/s of network capacity this year; it is migrating to multi-region Azure deployment, expected to be completed early next year. The principles behind the Git architecture changes are simplifying primitives, decoupling storage from compute, and reducing coordination. For ordinary users, these numbers are harder to understand than Fusion; for people blocked every day by CI and PR queues, they determine whether agent workflows can actually run.

GitHub's vision is to connect planning, coding, testing, and deployment into one continuous system. Models are one layer; review, permissions, cost, and infrastructure are all on the same bill. Fusion's cost-performance is worth trying, but don't just look at the scores in the keynote. Pick your messiest repository, have an agent open a PR, and then decide whether to hand over bigger tasks.

Original article: GitHub Universe 2026: Evolving every layer of the stack
X original post: GitHub on X