{"id":257637,"date":"2026-11-10T00:00:35","date_gmt":"2026-11-10T08:00:35","guid":{"rendered":"https:\/\/devblogs.microsoft.com\/visualstudio\/?p=257637"},"modified":"2026-11-10T09:56:31","modified_gmt":"2026-11-10T17:56:31","slug":"copilot-in-visual-studio-just-got-a-major-upgrade","status":"publish","type":"post","link":"https:\/\/devblogs.microsoft.com\/visualstudio\/copilot-in-visual-studio-just-got-a-major-upgrade\/","title":{"rendered":"Copilot in Visual Studio Just Got a Major Upgrade"},"content":{"rendered":"

Copilot Chat in Visual Studio just got a new engine. <\/strong>Starting with Visual Studio 2026 version 18.12, it runs on the Microsoft Copilot SDK<\/a> developed by GitHub, the same technology powering GitHub Copilot CLI, the GitHub Copilot app, and VS Code. It\u2019s on by default, and it\u2019s a major upgrade: same models, more tasks solved, new ways to work<\/strong>.<\/p>\n

We put the previous and new Copilot experiences through over one hundred coding tasks. The new experience came out ahead with every model we tested, across capability levels and providers, resolving up to 21 more tasks<\/strong>.<\/p>\n

The chat itself is better too: you get shorter, more focused answers, you can steer Copilot mid-task and see what each request costs.<\/p>\n

\"Copilot<\/p>\n

More Tasks Solved<\/h2>\n

To measure what matters to .NET developers, the .NET team built a benchmark from real issues in real C# projects, across .NET Framework and .NET Core, Azure and Aspire, and WinForms and desktop apps. We ran both Copilot chat experiences through 157 of its tasks. A task counts as resolved only when its failing tests pass without breaking any that already passed.<\/p>\n

\"Bar<\/p>\n

The new engine has also been tested on C++, Python, and a range of other languages.<\/p>\n

One example from our own backlog: a user reported that long answer choices in Copilot’s questions were getting cut off. We asked Copilot to help reproduce it, once in the previous experience and once in the new SDK-powered experience. The previous experience came back with steps for us to reproduce it by hand. The new one did something we didn\u2019t expect: it realized it was running inside the very chat window that had the bug, so it reproduced the problem right there.<\/p>\n

\"Side-by-side<\/p>\n

Same Models, Better Harness<\/h2>\n

On its own, a model reads text and writes text. It can’t open your solution, edit a file, or run your tests. The harness is everything around the model that turns text into real work: what the model sees, which tools it can use, what it remembers, and what it’s allowed to touch. Give the same model a better harness, and you get better results.<\/p>\n

Through the SDK, Copilot in Visual Studio now runs on the Microsoft Copilot harness<\/a>. It’s battle-tested at the scale of millions of users, and GitHub designed it to be fast and predictable<\/a>. Here’s what you’ll notice:<\/p>\n