Codex Users Hitting ‘At Capacity’ Errors Fell 80x in 24 Days

The share of Codex daily active users who hit a “model at capacity” error fell from 6.82% on 19 October to 0.085% on 12 November, according to a chart posted on X on Monday by Thibault Sottiaux, who works on Codex and ChatGPT at OpenAI, per his X profile. That is about 80 times lower in 24 days, by our arithmetic from the two values printed on OpenAI’s chart.

Put another way, about 1 in 15 Codex daily users had at least one turn end in a capacity error at the 19 October peak; on 12 November it was about 1 in 1,200, by our arithmetic. The post had about 702,000 views by mid-afternoon UTC.

What Sottiaux Posted

Sottiaux wrote at 05:55 UTC: “Codex team might have actually fixed the “model at capacity” errors that have been pestering you”. The attached chart is titled “Codex users hitting a capacity error” and covers 15 October to 12 November 2026, one point per day, in UTC.

The chart’s subtitle defines the metric: “Share of daily active users with at least one turn ending in a capacity error”. It is a share of users, not a count of errors or of requests, and the chart gives no number of users.

A month earlier, on 15 October, Sottiaux posted that the team had “landed a series of changes and more compute to very significantly decrease the occurrence of the disruptive” error, the message “Selected model is at capacity. Please try a different model”.

Share of Codex daily active users with at least one turn ending in a capacity error, 15 October to 12 November
Share of Codex daily active users with at least one turn ending in a capacity error, 15 October to 12 November 2026, re-drawn from the chart Thibault Sottiaux of OpenAI posted on X on 16 November. Only 6.82% (19 October) and 0.085% (12 November) are printed on OpenAI’s chart; the other daily points are our approximate reads. The roughly 80-times drop is our arithmetic.

Business Pill · Inference

An evergreen Business Engineer explainer on inference, the cost of running a trained AI model for every request. It teaches the general idea only and does not discuss Codex or OpenAI’s chart.

The key insight: As we read the chart, the metric counts users whose turn ended in an error, so OpenAI could lower it in two ways its posts name: more compute, and clients that wait and retry instead of failing. The six markers on the chart are all of the second kind, by their labels.

What Was Posted on X

Thibault Sottiaux, who works on Codex and ChatGPT at OpenAI per his X profile, posted this on 16 November 2026 at 05:55 UTC with the chart of Codex users hitting a capacity error. The piece’s 6.82% and 0.085% figures are the two values printed on that chart.

Sottiaux’s earlier post of 15 October 2026, saying the team had landed changes and more compute to reduce the “Selected model is at capacity” error. The piece quotes it for what OpenAI said it changed.

What the Chart Shows

The line is volatile through late October and early November. By our reading of the plotted points, 10 of the 22 days from 15 October to 5 November sit at or above about 3%, with spikes near 5% on 26 and 28 October. The line averages about 2.6% over those 22 days, by our arithmetic from approximate chart reads.

From 7 November the line flattens near zero. Every plotted day after the 6 November marker is below about 0.25%, and the six days from 7 to 12 November average about 0.05%, by our arithmetic from the same reads. Only 6.82% and 0.085% are printed on the chart; every other value here is read off the line and is approximate.

The Six Retry Markers

The chart carries six numbered markers under the heading “Retry-After release markers”. Retry-After is the standard web signal a server sends to tell a client how long to wait before trying again. By their labels, all six concern when and how Codex and OpenAI’s Responses service delay and retry requests.

The first, on 15 October, reads “Desktop release carries WebSocket retry fix”. On 22 October, “Responses adds a delay on ordinary overloads”; on 23 October, “Service-tier overloads get a retry time”; on 24 October, “Codex forwards retry and stop instructions”.

The last two are dated 30 October, “Responses adds a delay for NO_MORE_RETRY”, and 6 November, “Responses keeps the delay in WebSocket errors”. The near-zero stretch on the chart starts the day after that sixth marker.

Users had been reporting the error for weeks. An issue opened on 8 September in OpenAI’s public Codex repository on GitHub collects reports of the “Selected model is at capacity” error and was still open on Monday. A user issue on 26 October said Codex had stopped retrying on its own after the error: before, it “would display a retry countdown, wait, and automatically attempt to resume the task without requiring user intervention”.

How It Fits the Capacity Squeeze

Capacity is now shaping AI products across labs. Our DeepSeek piece covers its free chat being rationed, our Claude Ultra piece measures what Anthropic’s top plan allows a month, and our piece on OpenAI’s 2027 capacity covers a report on its compute plans.

OpenAI’s chart tracks one thing a user feels directly: whether a turn ends in an error. By the chart’s own definition, a turn that waits and then succeeds is not counted, so the metric tracks the experience of capacity limits rather than GPU supply.

The Structural Read

As we read the numbers, at the 19 October peak about 1 in 15 Codex daily users had a turn end in a capacity error; on 12 November it was about 1 in 1,200, by our arithmetic from the two values printed on OpenAI’s chart.

As we read the timeline, the line stayed volatile through the first five retry markers and flattened only after the sixth, on 6 November; the chart itself does not attribute the drop to any one change.

Thibault Sottiaux, Codex and ChatGPT at OpenAI, on X, 15 October 2026

“When it does happen, it should also eventually just continue by itself and not just block on it”

Three Implications

FOR CODEX USERS On 12 November the chart shows 0.085% of daily users with a turn ending in a capacity error, down from 6.82% on 19 October.

FOR DEVELOPERS Three of the six markers name OpenAI’s Responses service, per the chart’s labels; the other three name a desktop release, service-tier overloads and Codex itself.

FOR RIVAL LABS The chart puts a public number on one reliability measure for a coding agent, a figure our related pieces on DeepSeek and Claude did not have for those products.

The Business Engineer Lens

This story maps onto the Business Engineer framework The State of the Inference Economy.

The framework’s thesis: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query”.

As we read it, a capacity error is the moment that marginal cost is refused: the chart shows OpenAI cutting how often a Codex turn ends that way, from 6.82% of daily users to 0.085%. The framework predates the chart and does not discuss Codex.

What Is Not Established

The chart does not say how much of the drop came from added compute and how much from retry handling; the October post names both. It gives no count of users, errors or requests, does not cover days after 12 November, and does not break the figure down by model, plan or surface. Sottiaux’s own wording is hedged: the team “might have actually fixed” the errors.

The 80-times figure compares a single peak day with a single recent day; on the averages we read off the chart the drop is about 50 times, by our arithmetic. OpenAI was not contacted for this piece.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The State of the Inference Economy

The Business Engineer read on why inference, not training, now sets the economics of AI.

Read the premium analysis →Get 50% off →

The Bottom Line

OpenAI’s chart shows the share of Codex daily users hitting a capacity error at 0.085% on 12 November, down from 6.82% on 19 October, about 80 times lower in 24 days by our arithmetic. All six release markers on the chart concern retry timing, per their labels.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. The chart, its labels and the quoted posts are from Thibault Sottiaux’s posts on X of 16 November and 15 October 2026, read in full. The chart’s printed values are 6.82% and 0.085%; every other daily value is our approximate read of the plotted line, and the re-drawn chart is ours. User reports are from two issues in OpenAI’s public Codex repository on GitHub. The 80-times ratio, the 1-in-15 and 1-in-1,200 conversions and the averages are our arithmetic. OpenAI was not contacted for this piece.

Sources: Thibault Sottiaux (OpenAI) on X, 16 Nov 2026, with chart · Thibault Sottiaux on X, 15 Oct 2026 · openai/codex GitHub issue #43738 · openai/codex GitHub issue #57040

Scroll to Top
FourWeekMBA