OpenAI has introduced Pando, a new Git backend that now handles all development in its monorepo, meaning a single code repository. The Git client and protocol remain unchanged: engineers and Codex agents continue to run git push and git fetch commands as before, only the server has changed. According to the company, the median latency for code pushes (push) has dropped by nearly 80%.
- OpenAI introduces Pando, a new Git server backend that handles its monorepo.
- The median push latency drops by nearly 80%, according to OpenAI.
- Pando handles each branch separately and stores each pack as an immutable blob.
- The figures are self-reported, and Pando remains internal and closed source.
Why OpenAI changed its backend
Codex agents have multiplied parallel work. According to the commit graph published by OpenAI, commits on the master branch increased by 1,378% in Q3 2026 compared to Q1 2024, with an estimate for Q4 2026 reaching +2,442%. Each push now triggers many code fetches by continuous integration tools. By late 2025, fetches and pushes were slowing down, merges were failing, and reliability sometimes dropped below 90%. The company concluded the bottleneck was in how it served Git, not Git itself.
How Pando works
According to the engineering post published by OpenAI, Pando discards two assumptions that most Git servers consider unavoidable. The first concerns repository-wide consistency: classic servers order every reference update for the entire repository, so unrelated pushes must wait. Pando adopts causal consistency, where each branch is handled separately and clients only conflict when pushing the same reference. The trade-off is that a reader consulting multiple branches may see them at different points in time.
The second assumption concerns generating a separate pack for each client. Standard servers create a new one for each fetch, while Pando stores each sent pack as an immutable blob and records a “reference version” in an extensible metadata store. A push requires a constant number of metadata reads and writes, while a fetch requires a number proportional to the logarithm of the version gap to fill. To limit push size, only the main branch is advertised as a base for new branches. In OpenAI’s monorepo, the median pack weighs less than 20 KiB.
Packs are then distributed via packfile-uri, an existing Git protocol feature, and a delivery service called Express caches them by region, with a success rate above 99%. Each push creates an “L0” pack that background processes merge into larger and larger packs, up to level L4 for the current monorepo. This mechanism replaces Git’s traditional garbage collection.
Self-reported figures
OpenAI states that Pando handles nearly a petabyte of Git data per day using only a few CPU cores, and that it served over 57 TB of packs in a single hour, with over 99% from the local disk cache. On standard cloud VMs, clones and fetches saturate 10 Gbit/s links. The company also conducted tests with several thousand simultaneous pushers, with no degradation reported. All these figures are self-reported, and no independent benchmark has been published yet.
Pando is not open source, and OpenAI says it is exploring ways to share the technology. Reads spanning multiple branches do not have a single snapshot of the repository at a given moment, and repeated pushes of stacked branches return all content outside the main branch. Patches to the Git client have been upstreamed, while the backend remains internal for now.
