OpenAI describes Pando as the Git-compatible backend powering development in its monorepo. In an engineering post published November 2, 2026, the company reported that Pando reduced median production push latency by nearly 80%. OpenAI says the system began serving production traffic in April 2026, after about 10 weeks of development. OpenAI’s engineering post explains how it handles repository updates and fetches.
How Pando handles pushes and fetches
Pando validates incoming Git packfiles as it streams them to blob storage, while a separate metadata store tracks reference versions. A reference is a named pointer such as a branch. This design lets Pando record branch updates without unpacking every incoming pack into a local repository.
For a fetch, Pando uses the client’s history and requested commits to select stored packs that cover the missing objects. It returns packfile URLs, and Express delivers the data through regional caches. Background compaction groups packs into larger covers, reaching L4 in OpenAI’s monorepo. Reusing packs avoids generating a custom pack for every reader.
There is a tradeoff when developers push to feature branches. Pando generally advertises a designated trunk branch as the base for a new feature branch, keeping branch advertisements independent of the repository’s growing branch count and making packs easier to reuse. Repeated pushes to a feature branch or stacked branches can resend objects outside that trunk base. In OpenAI’s monorepo, packs produced under this approach had a median size below 20 KiB; the 99th-percentile size was below 3 MiB.
What parallel branch updates mean for readers
Pando orders updates to each individual reference serializably, while allowing updates to unrelated references to proceed in parallel. That means a read spanning several branches is not guaranteed to show them all at one point in time: different branches can reflect different moments during concurrent updates.
How OpenAI’s performance figures are scoped
OpenAI reports nearly one petabyte of Git data delivered per day. At a peak, it says Pando served more than 57 TB of Git packs in one hour, with more than 99% of those bytes coming from local disk cache. The company also says validated clone and fetch operations saturated 10 Gbps links on standard cloud-provider virtual machines. These are figures for OpenAI’s system and workload.
OpenAI’s interactive comparison of Pando with other Git architecture models is labeled a synthetic teaching model, not a production benchmark. It explains modeled tradeoffs; it is separate from the company’s reported production results.
What OpenAI says about sharing Pando
OpenAI describes Pando as an internal system and says it is exploring ways to share the technology that enables it. The company has not set a release date or format for Pando itself. Separately, OpenAI says some Git improvements from the work are already upstream and others are in its public repository.
