Pando
OpenAI describes Pando, an internal Git backend built with Codex in 10 weeks that cut push latency by almost 80% and serves nearly 1 PB of Git data a day.
Pando keeps Git protocol compatibility but separates write coordination from read optimization, using packfile reuse, to cope with the heavy concurrent load from AI agents such as Codex. OpenAI says it was built with Codex in 10 weeks. It remains internal, and OpenAI says it is evaluating how to share it.
- Date
- Monday 2 November 2026
- Lab
- OpenAI
- Kind
- infra
- Access
- paper only
Figures
| Measure | Value | Measured by |
|---|---|---|
| Push latency reduction | almost 80% | company |
| Git data served per day | nearly 1 PB | company |
| Peak hourly throughput | over 57 TB | company |
| Pack retrieval cache hit rate | 99%+ | company |
Pando is internal to OpenAI and not available to use. Only the engineering post, apparently titled 'Scaling Git at OpenAI', is public, and OpenAI also published modified Git client code on GitHub.
Sources
This record was partly confirmed: some claims could not be checked on 4 November 2026. How we check