Putting o1-preview to Work Inside Copilot

OpenAI’s newly released o1 model series is built around advanced reasoning, and GitHub has been testing how that capability transfers to Copilot. The initial experiments focus on code analysis and optimization, where o1-preview’s ability to think through a problem before responding lets Copilot break complex tasks into structured steps. Two demos illustrate where the model fits in daily development work.

Optimizing a Tokenizer with the Built-In Optimize Command

The first scenario involved a real performance issue faced by the VS Code team: Copilot Chat must repeatedly tokenize large data volumes while assembling prompts. A new built-in Optimize chat command supplies rich context from the editor, including imports, tests, and performance profiles.

GitHub used the command to refine a byte pair encoder inside Copilot Chat’s tokenizer library—meaning AI was used to optimize a core building block of AI development. The test showed o1-preview’s reasoning allows a deeper grasp of code constraints and edge cases, yielding a more efficient result in one shot. GPT-4o, in contrast, stuck to obvious optimizations and needed developer guidance to reach more complex approaches. The model also handled the math side of the task, calculating benchmark results from raw terminal output and summarizing them without extra prompting.

Debugging a Focus Management Bug on GitHub.com

In the second demo, o1-preview diagnosed and fixed a performance bug in GitHub’s own focus management code. The issue surfaced when adding a folder tree to the file view on GitHub.com—with 1,000 elements under management, the code stalled and crashed the browser. The bug had previously taken a software engineer a few hours to solve; o1-preview produced the same solution in minutes. The change improved runtime from more than 1,000ms to roughly 16ms.

The key difference was in response style: o1-preview gave deliberate, line-by-line recommendations that made the problem easy to isolate. GPT-4o, given the same prompt, was more likely to return a blob of code rather than a structured breakdown of the fix.

o1-Preview and o1-Mini Arrive on GitHub Models

GitHub is also making the o1 series available through GitHub Models. Both o1-preview and o1-mini—a smaller, faster model priced 80% lower—land in the marketplace shortly. Because the models are still in preview, access requires signing up for Azure AI for early access.