# Pando

> Source: https://aiwiki.ai/wiki/pando
> Updated: 2026-11-04
> Categories: AI Infrastructure, Developer Tools, OpenAI, Software Development
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Pando." aiwiki.ai, 4 Nov 2026. https://aiwiki.ai/wiki/pando
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Pando** is [OpenAI's](https://aiwiki.ai/wiki/openai) internal Git backend for its monorepo. It supports concurrent branch updates and cached distribution of repository history while retaining Git workflows for engineers and [coding agents](https://aiwiki.ai/wiki/ai_coding_agent). OpenAI described it publicly on November 2, 2026.[1]

## Git objects and references

Git stores file contents as blobs, directory structures as trees, and history as commits linked to trees and parent commits. References identify branch tips. Object identifiers let unchanged content be reused across revisions.[3] These distinctions matter when serving a repository: transferring file-history objects and publishing a branch's new tip are separate operations.

Git's fetch protocol lets a client describe commits it wants and commits it already has. The server can then supply the missing history. A response can include both a packfile and links to additional packfiles, rather than transferring every object through one connection.[4]

## Architecture and consistency

Pando separates immutable packfiles in blob storage from reference versions and their dependencies in a metadata store.[1]

| Operation | Pando's documented design |
| --- | --- |
| Push | Stream and validate a pack, then atomically update reference metadata |
| Fetch | Select reusable packs from the client's existing history |
| Delivery | Serve packs through regional Express caches |
| Compaction | Combine small packs into larger, reusable covers |

Updates to one reference are serializable; unrelated branches can advance independently. Reads spanning several branches do not provide a repository-wide point-in-time snapshot. This is the design's causal-consistency tradeoff.[1]

### Packfile URI transport

Git protocol version 2 includes the `packfile-uris` capability. A client advertises acceptable transfer protocols, and a server can respond with packfile hashes and download locations. The client downloads the referenced packs before checking object connectivity. This protocol separates negotiation from bulk object delivery.[4][5]

Pando uses that mechanism for reusable packs and sends an empty inline packfile. Its default new-branch advertisement uses trunk as the base, trading some redundant uploads for reusable downloads.[1]

## Reported production results

OpenAI's developer announcement reported roughly 80% lower push latency and nearly one petabyte of Git data delivered per day. These are company-reported production figures, not an independent benchmark.[2] The engineering article separately labels its interactive architecture comparisons as synthetic teaching simulations.[1]

## Public Git development

OpenAI maintains a public fork of Git, whose default displayed branch is `codex`. This repository contains Git client and related implementation work; it should be distinguished from the internal Pando service.[6]

The company's public patch tracker separates changes included in its build from development branches. It records work on parallel packfile-URI downloads, progress reporting, download recovery, and reuse of packs already downloaded. It also lists `explicit-haves` changes for push negotiation. A development-branch listing does not mean the change is included in released upstream Git.[7]

The standard Git documentation describes packfile URIs as a mechanism through which a server can offload transfer to other delivery infrastructure. OpenAI's extensions build on that existing protocol rather than introducing a replacement file-history format.[4][5][7]

## Availability

The announcement describes internal deployment and explores future sharing; it does not announce a public Pando service or backend release.[1]

## References

1. Aaron Friel, Taylor Blau and Ted Nyman. [Scaling Git at OpenAI](https://openai.com/index/scaling-git-at-openai/). OpenAI engineering article, November 2, 2026.
2. OpenAI Developers. [Pando production announcement](https://x.com/OpenAIDevs/status/2117408792346509663). November 3, 2026.
3. Scott Chacon and Ben Straub. [Git Internals: Git Objects](https://git-scm.com/book/en/v2/Git-Internals-Git-Objects). Pro Git, accessed November 4, 2026.
4. Git project. [Protocol version 2](https://git-scm.com/docs/protocol-v2). Fetch negotiation and packfile-URI response sections, accessed November 4, 2026.
5. Git project. [Packfile URI documentation](https://git-scm.com/docs/packfile-uri). Protocol, server and client design, accessed November 4, 2026.
6. OpenAI. [Public Git fork](https://github.com/openai/git). Repository identity and code, accessed November 4, 2026.
7. OpenAI. [Git patches and notes](https://oai-git.openai.chatgpt.site/). Public build and development-branch tracker, accessed November 4, 2026.