# Codex Capacity Errors Fell 80x: What OpenAI Changed

> OpenAI's Codex lead says capacity errors fell from 6.82% of daily users on Oct 19 to 0.085% on Nov 12, after six retry fixes. What his chart shows.

- Author: Nitish Garg, Founder & CEO, CellCog
- Published: 2026-11-16
- Canonical (HTML): https://cellcog.ai/blog/codex-capacity-errors-fixed/
- Section: Guides / Choosing a platform
- Publisher: CellCog (https://cellcog.ai), the AI employee platform. Blog index for agents: https://cellcog.ai/blog/llms.txt

## Key points

- On November 16, 2026, Codex lead Thibault Sottiaux posted that the Codex team might have fixed the 'model at capacity' errors users kept hitting.
- His chart tracks the share of daily active Codex users with at least one turn ending in a capacity error, from October 15 to November 12.
- The share peaked at 6.82% on October 19 and stood at 0.085% on November 12, about 80 times lower.
- The chart marks six releases between October 15 and November 6 that changed how Codex and the Responses API handle retries and overload delays.
- After the sixth, on November 6, the line stays near zero for the rest of the chart.
- OpenAI has published no write-up; the chart is one post, and it shows a share of users, not counts or causes.

## At a glance

- **Did OpenAI fix Codex's capacity errors?** Its Codex lead says the team 'might have actually fixed' them. His chart shows the share of daily users hitting one fell from 6.82% at the October 19 peak to 0.085% on November 12.
- **What changed?** Six releases, per the chart's markers, mostly about retries: a WebSocket retry fix, delays on overloads, retry times for service-tier overloads, and Codex passing retry and stop instructions through.
- **Is the chart official?** It was posted by Thibault Sottiaux, who leads Codex at OpenAI, on his own X account. OpenAI has not published it on its blog or changelog.

**The share of Codex users who hit a capacity error fell from 6.82% to 0.085% in under four weeks, according to a chart posted by OpenAI's Codex lead.** Thibault Sottiaux [posted on X](https://x.com/thsottiaux/status/2122206990101536852) at 05:55 UTC on November 16, 2026 that the Codex team "might have actually fixed" the errors. The post had about 686,000 views when we read it at 14:02 UTC.

The chart is titled "Codex users hitting a capacity error" and covers October 15 to November 12, in UTC. It measures one thing: the "Share of daily active users with at least one turn ending in a capacity error". It is a share of people, not a count of errors, and it is one post on a personal account, not an OpenAI write-up.

## The two numbers

*Table: Codex users hitting a capacity error, share of daily active users (per the chart)*

| Day | Share of daily users | Note |
|---|---|---|
| October 19, 2026 | 6.82% | The peak, labelled on the chart |
| November 12, 2026 | 0.085% | The last day charted, labelled |

By our arithmetic, 6.82% to 0.085% is about 80 times lower. On a day at the peak, roughly one daily Codex user in fifteen had at least one turn end in a capacity error; on November 12 it was under one in a thousand. Between the peak and the end, the line swings: it falls below 1% around October 24, climbs back near 5% in late October, drops again after October 30, and spikes once more in early November before settling near zero from November 7.

## Six releases, all about retries

The chart's legend is headed "Retry-After release markers". Each marker is a release that changed how Codex or OpenAI's Responses API handles an overload, not one that added servers.

*Table: The six release markers on Sottiaux's chart*

| Marker | Date | What the chart says shipped |
|---|---|---|
| 1 | October 15 | Desktop release carries WebSocket retry fix |
| 2 | October 22 | Responses adds a delay on ordinary overloads |
| 3 | October 23 | Service-tier overloads get a retry time |
| 4 | October 24 | Codex forwards retry and stop instructions |
| 5 | October 30 | Responses adds a delay for NO_MORE_RETRY |
| 6 | November 6 | Responses keeps the delay in WebSocket errors |

Read together, the markers describe one idea: when the model is busy, tell the client how long to wait, and have Codex wait and try again instead of ending the turn. That is our reading of six one-line labels; Sottiaux did not explain the mechanism in the post. The drop after the sixth marker, on November 6, is the clearest on the chart: the line goes from about 2% to near zero within two days and stays there through November 12.

## What the chart does not say

- **Whether capacity grew.** Every marker is a retry or delay change. A user whose request is retried may wait longer instead of seeing an error. The chart does not show wait times.
- **Days after November 12.** The post went up on November 16; the chart stops four days earlier.
- **Counts or plans.** It shows a share of daily active users, with no user counts and no split by plan, surface or model.
- **An official record.** We found nothing on OpenAI's [API changelog](https://developers.openai.com/api/docs/changelog), read at 14:00 UTC on November 16, newer than November 5. The chart is Sottiaux's own post.

Replies to the post split between users who say the errors stopped and users who say others simply moved to other tools. One [reply](https://x.com/AnikaCodes/status/2122263487665897485) with about 1,000 views argues the drops line up with the retry fixes, not with people leaving. Neither side gives data, so we leave it there.

## Where CellCog sits

Our conflict, declared: we build CellCog, a platform where a business hires AI employees, and our agents run on Anthropic's Claude models, not on Codex. Codex is a tool many of our readers use beside us. For earlier Codex changes, see our posts on [Codex persistent mode](https://cellcog.ai/blog/codex-persistent-mode/) and the [shared ChatGPT and Codex usage limit](https://cellcog.ai/blog/chatgpt-codex-shared-usage-limit/).

## Sources

- Thibault Sottiaux on [X](https://x.com/thsottiaux/status/2122206990101536852), November 16, 2026, 05:55 UTC, with the chart "Codex users hitting a capacity error" ([image](https://pbs.twimg.com/media/HXOX68AaYAAlezW.jpg); labels transcribed by us).
- A [reply](https://x.com/AnikaCodes/status/2122263487665897485) by @AnikaCodes, November 16, 2026, about 1,000 views when read.
- OpenAI's [API changelog](https://developers.openai.com/api/docs/changelog), read November 16, 2026 at 14:00 UTC; newest entry November 5.

## FAQ

**What is a 'model at capacity' error in Codex?**

A turn that ends because the model is overloaded. The chart counts a user once per day if at least one of their turns ended that way.

**How big is the drop?**

From 6.82% of daily active users on October 19 to 0.085% on November 12, about 80 times lower by our arithmetic.

**Does this mean Codex has more capacity?**

The chart does not say. Its markers are retry and delay changes, which suggests overloads are now retried instead of shown to the user, but OpenAI has not explained it.

**What about November 13 to 16?**

The chart stops at November 12. Days after that are not shown.

**Where does CellCog fit?**

We build AI employees for any role, and our agents run on Anthropic's Claude models, not Codex. An AI software engineer on CellCog reads changes like this for your team.

## Related

- [Codex Persistent Mode: OpenAI's Always-On Agent, What We Actually Know](https://cellcog.ai/blog/codex-persistent-mode/index.md)
- [ChatGPT and Codex Will Share One Usage Limit in December](https://cellcog.ai/blog/chatgpt-codex-shared-usage-limit/index.md)
- [Codex Wrecked a Server After a Reboot Request](https://cellcog.ai/blog/codex-server-reboot-incident/index.md)
- [Codex Auto-Review Is Now Free: What It Blocks, How to Enable](https://cellcog.ai/blog/codex-auto-review/index.md)

## The AI employee for this read

[AI Software Engineer](https://cellcog.ai/ai-employees/ai-software-engineer): I built this page. For what it covers, hire an engineer: it works in your repo behind an approval gate, so nothing reaches your world unclassified.

---

Markdown alternate of https://cellcog.ai/blog/codex-capacity-errors-fixed/. Try CellCog free, no credit card needed: https://cellcog.ai/signup
