GitHub says Copilot is 77% cheaper, and the comments are asking who presses the merge button

Satya Nadella posted on X saying GitHub is redefining the SDLC for the agent era. On the same day, the GitHub Universe 2026 blog laid out the details, written by Jay Parikh under the title “Evolving every layer of the stack.”

In the blog, Jay Parikh writes that the SDLC is no longer a linear path of plan, code, test, deploy. Code changes, a colleague’s request, production metrics—work can be triggered from anywhere. Developers set direction, make decisions, and own delivery; agents handle implementation and validation.

The most concrete numbers come from Copilot Fusion. GitHub’s claim is frontier-model-level quality, 40% faster, and 77% lower cost. On the Artificial Analysis Coding Agent Index, Fusion Max is two points behind Opus 5.5 Max, at half the cost, and spends 40% less time per task. Mark asks in the comments: Is Fusion routing between multiple models, or is it a new model? According to GitHub’s docs, Fusion evaluates what a task needs and decides how to execute it across dimensions such as reasoning, code generation, debugging, and tool calling, and it can span one or more models. So Fusion does routing and orchestration; it did not launch a new model.

Copilot Projects is another highlight. It lets you run agent tasks across multiple repositories, fully managed, with parallel sessions. Anand Butani says real work already spans multiple repositories, and being able to run agents in one place is exactly what’s been missing. Mr. Clicker asks more directly: a loop across multiple repositories sounds great, until two agents step on each other’s changes. Who wins that merge? Aleksandar Janca adds: with agents running across multiple repositories at the same time, one bad change can land in several places at once, and he wants a human to approve every merge.

Those concerns are grounded. In a reply, dipeshpatel says almost anything can be vibe coded now, but without solid engineering practices, it’s hard to guarantee quality for enterprise-grade products. More agents just means more code that someone has to review. The Lunar Misfit’s joke is sharper: multiple agents across multiple repositories, all asking whether GitHub is down.

The Foundry piece is also worth a look. GitHub’s unified Copilot harness has entered Foundry and become a managed service for running long-running agents that can reason, call tools, and do work. The harness can be understood as the runtime that drives agents; you no longer have to build it yourself.

A few smaller updates are useful too. Session continuity lets you follow the same cloud agent session across the GitHub app, VS Code, CLI, and web. Start on your phone, switch to VS Code at your desk, check in from github.com at lunch, and the agent keeps working in the cloud. Copilot automations let you build AI workflows in a repo, run them in the cloud, trigger them manually or on events or schedules, and now learn from results to suggest improvements for you to review. Copilot code review Max has the highest F1 score on high-severity issues in ReviewBench; Lite and Balanced handle everyday quick reviews, while Max is for deeper inspection.

There’s a commercial change. GitHub Enterprise customers can now use all Copilot features without buying a separate Copilot subscription, including the app, code review, automations, and more, paying by AI credits used. Existing Copilot Business and Enterprise customers can keep their current plan or migrate at any time.

Infrastructure numbers: this year GitHub added 4 million CPU cores, 160 PB of high-speed storage, and 50 Tb/s of network. It is migrating toward multi-region Azure deployment, expected to complete early next year. The Git architecture is also changing, with principles of simplifying primitives, decoupling storage and compute, and reducing coordination. Jay Parikh writes in the blog: all the software written so far may add up to less than 1% of the software that will be written in the next five years.

In the comments, Jon Cheney focuses on the 77% cost reduction. He says the companies he talks to are no longer mostly stuck on quality; once they get past the pilot team, the bill is the problem. That observation is closer to most teams’ reality than the launch event itself.

Tory says GitHub’s moat is shifting from code hosting to workflow gravity. If Fusion can lower inference costs and Projects can coordinate multiple agents across repositories, then with every additional agent, GitHub moves one step closer to the center of the software-building process. That take is more accurate than simply saying Copilot is a better coding assistant.

There are also tangents. Someone asks whether Microsoft actually makes models, someone asks what happens to Azure DevOps, and someone just starts ranting. Comments section, you know.

Finally, there’s a funny video where someone pairs a World of Warcraft character with the line “You didn’t MERGE my PULL request.”