OpenAI’s Pando cuts Git push latency by about 80% for its monorepo

OpenAI has built its own Git backend. It is called Pando, and it reportedly makes pushing code to the company’s monorepo about 80% faster.

That works out to a nearly fivefold drop in median push latency.

The bigger story sits behind the speed number. OpenAI says its repository was buckling under a surge of commits from Codex, its AI coding agent.

What Pando actually does

OpenAI announced Pando on November 2, 2026. According to the research findings, the system went into production in April 2026 after about 10 weeks of development.

The build team was unusually small. One engineer led the project, with substantial help from teams of Codex agents.

Pando tackles two separate scaling headaches. The first is many people writing to different branches at the same time. The second is heavy read traffic, such as fetches and clones.

Pando allows parallel updates to unrelated branches while keeping the order of changes consistent within each branch. The technical term is causal consistency per branch, and in practical terms it means edits on a single branch still land in the sequence they were made, even as everything else runs in parallel.

Crucially, developers do not have to change how they use Git. Pando keeps full compatibility with the Git protocol and existing clients. OpenAI has also contributed work back to the Git project itself.

The numbers behind the backend

The throughput figures are hefty. Pando serves nearly one petabyte of Git data per day, and it does so with just a handful of CPU cores.

At its recorded peak, the system pushed out more than 57 TB of Git pack data in a single hour. Over 99% of that came straight from local disk cache.

Clients can also pull data quickly. The research notes that clone and fetch operations can use up to 10 Gbps of bandwidth on standard cloud virtual machines.

Why OpenAI needed a new Git server

The pressure came from Codex. According to the research findings, Codex-driven commits rose 1,378% from Q1 2024 to Q3 2026.

Forecasts point higher still, with overall growth suggested at 2,400% by the end of 2026.

Traditional Git server architectures reportedly struggled with this load. The research findings say reliability often dipped below 90% under the rising demand.

Pando also improves horizontal scaling. OpenAI can add more machines to handle more traffic rather than trying to make a single server ever more powerful.

What this means for developers and the industry

There is also a signal about who builds infrastructure now. One engineer plus Codex agents shipped a production Git backend in about 10 weeks.