OpenAI's In-House Code Store, Pando, Cuts Push Times by 80%
Built to absorb the flood of development work generated by Codex, the internal system stays compatible with Git and went from kickoff to production in 10 weeks.
On November 2 local time, OpenAI published the design and operational results of Pando, its code storage system. The system was built to absorb the large volumes of development work pouring out of Codex, the company’s coding AI. OpenAI says the time it takes to “push” code changes to the server has dropped by roughly 80% at the median in its live production environment.
Pando is compatible with Git, the tool used to manage code change history. Today it backs the repository where OpenAI keeps code for many of its projects in one place. The company says it began building Pando with Codex and put it into production 10 weeks later, with the system handling real production traffic starting in April 2026.
Why it stays fast under heavy load
Pando processes changes arriving from different branches at the same time. A branch is a separate workspace that splits off from the same codebase. The key idea is that even when code arrives from many workspaces at once, the work proceeds in parallel. OpenAI says it designed the system to handle that kind of large-scale concurrency while preserving compatibility with existing Git.
On the download side, Pando reuses bundles of data it has already assembled to serve requests from multiple users. These bundles are called Git packfiles. The system also caches data regionally and merges packfiles in the background. OpenAI says this approach raised both download speed and the volume of data it can serve at once.
According to OpenAI, Pando delivers about 1 petabyte of Git data per day. In one instance it shipped more than 57 terabytes of packfiles in a single hour. The share of requested data served directly from cache exceeded 99%. When cloning a repository or fetching changes, the company says it measured throughput that fully saturated a 10Gbps link.
The constraints developers should know
There are limits on operations that read many branches at once. Pando does not guarantee a read of the exact same point-in-time state across the entire repository. When pushing code to the server, data that has already been sent may be transmitted again. When that happens, the amount of network traffic used on the client side goes up.
How AI was used to build and verify it
OpenAI says it used GPT-6 Astra to rigorously specify Pando’s behavior and to fix bugs. The company says Daybreak-capable models are used to verify the security of both Git and Pando. In other words, it disclosed not only that it built the system with Codex but also how it verified and repaired that system using other AI models.
Pando is currently deployed inside OpenAI’s internal development environment. The company says it is weighing ways to share the underlying technology with users, but it did not offer pricing or a timeline for an external release. What form access to this technology might take for external developers remains to be seen.