While OpenAI did not release a formal root-cause analysis, the fix described by Codex lead Thibault Sottiaux provides a strong clue. He confirmed that the resolution involved "resetting the Codex rate limits across all plans" within a 24-hour window . This indicates that the problem was not a lack of physical compute, but rather that internal rate-limit ceilings were set too low for the sudden surge in demand, causing the system to reject valid requests as if it were full.

A smaller, related incident on June 11 had already signaled trouble, logging "elevated error rates for GPT 5.5 in Codex" . For the users paying the bill, this was the breaking point of building pressure, not an isolated event.

The Incident Timeline: A Three-Hour Scramble

The disruption had a long tail and a short, intense peak. External monitoring services first detected issues late on the evening of June 15, around 10:12-10:16 p.m. ET . However, the incident wasn't formally acknowledged until the next morning, leaving developers who rely on Codex for overnight tasks and early-morning work in a blind spot .

Once the internal alarms went off, the response was swift:

• 6:32 a.m. ET — OpenAI's status page marked the incident as "Identified" .
• ~7:39 a.m. ET — Vaibhav Srivastav publicly acknowledged the error on X, directing users to the status page .
• ~7:40 a.m. ET — Codex lead Thibault Sottiaux posted: "We are aware that some Codex users are experiencing high error rates with 'model at capacity' and are working to bring things back to being stable" .
• 9:32 a.m. ET — A mitigation was applied and monitoring began .
• 9:48 a.m. ET — The status flipped to green: "All impacted services have now fully recovered" .

The outage affected Codex across its entire surface area, including the CLI, VS Code extension, and Desktop app . While the official incident clock showed about 3 hours of trouble, the practical disruption for users spanned a much longer window.

Developer Frustration Reaches a Boiling Point

For professional and Pro-tier subscribers, this wasn't a minor inconvenience—it was an active threat to their productivity. The reaction on the OpenAI Developer Community and X was visceral.

The core complaint wasn't just that the service was down, but how it failed. Users reported that Codex sessions would "exit mid-stream without saving state," forcing them to manually reconstruct lost context and redo work . A user in the UK summarized the sentiment vividly: "Making it impossible to work as you don’t know at what point Codex exits and it is going over and over the same thing finding what was done and what wasn’t. Totally unacceptable" .

The generic error message itself was a major source of anger. The advice to "try a different model" provided no actionable guidance when the primary model was unavailable and users had no way of knowing whether to retry, downgrade their reasoning effort, or simply wait .

Trust in OpenAI's communication was also damaged. Several users pointed out a gap between when the problems actually began—by community reports and personal experience—and when the official status page clock started ticking, a discrepancy that makes incident transparency feel unreliable .

Amid the frustration, a slice of developer gallows humor emerged. Influencer Matthew Berman created willcodexquotareset.com, a site that jokingly displayed a "94% chance of a Codex quota reset in the next 48 hours" . Digg’s sentiment analysis of the conversation around the event showed a split: 63.8% positive, with many thanking OpenAI for the quick fix, but a significant 36.2% negative, with users calling the service's reliability into question after a string of repeated outages .

A Troubled Year: Codex Reliability in 2026

The June 15-16 incident is not a one-off. It is the most visible spike in a year of recurring Codex degradation that began in earnest in early May 2026. A pattern of GPT-5.5 capacity saturation and rate-limit mismatches has surfaced repeatedly.

A timeline of major Codex events in 2026 shows a platform under persistent stress:

• April 23 — The "Selected model is at capacity" error is first widely reported on the Community forum, with users noting mid-session failures and no state preservation .
• May 7+ — A cluster of Codex latency and reconnect-loop degradations begins, with measurable worsening in public response times .
• May 12 — Following a "2x capacity" rollout, Codex App stream disconnects become common, ending with websocket closure errors .
• May 13 — GPT-5.5 experiences high error rates, disproportionately hitting free-tier users .
• May 14 — Codex Cloud and Code Review services suffer a high failure rate for 32 minutes .
• May 29 — A full-stack OpenAI outage brings down ChatGPT, the API, DALL-E, Codex, and Sora simultaneously .
• June 3 — Elevated error rates hit Codex, ChatGPT, and the Responses API for about 6.2 hours .
• June 11 — Elevated errors for GPT-5.5 in Codex appear in the evening, setting the stage for the main event .
• June 15-16 — The "Selected Model is at Capacity" error becomes a full-blown outage lasting roughly 3 hours .

The common thread is clear: GPT-5.5 demand keeps slamming into configured ceilings, whether from rate limits, reasoning-effort overload, or broader infrastructure strain. The June 16 fix was a rate-limit reset, a remedy that treated the symptom—the ceiling being hit—rather than the underlying mismatch between capacity and the model's popularity. Without a deeper infrastructure scaling solution, the error remains a likely returnee as more developers adopt Codex for high-intensity coding tasks.

While OpenAI did not release a formal root-cause analysis, the fix described by Codex lead Thibault Sottiaux provides a strong clue. He confirmed that the resolution involved "resetting the Codex rate limits across all plans" within a 24-hour window . This indicates that the problem was not a lack of physical compute, but rather that internal rate-limit ceilings were set too low for the sudden surge in demand, causing the system to reject valid requests as if it were full.

A smaller, related incident on June 11 had already signaled trouble, logging "elevated error rates for GPT 5.5 in Codex" . For the users paying the bill, this was the breaking point of building pressure, not an isolated event.

The Incident Timeline: A Three-Hour Scramble

The disruption had a long tail and a short, intense peak. External monitoring services first detected issues late on the evening of June 15, around 10:12-10:16 p.m. ET . However, the incident wasn't formally acknowledged until the next morning, leaving developers who rely on Codex for overnight tasks and early-morning work in a blind spot .

Once the internal alarms went off, the response was swift:

• 6:32 a.m. ET — OpenAI's status page marked the incident as "Identified" .
• ~7:39 a.m. ET — Vaibhav Srivastav publicly acknowledged the error on X, directing users to the status page .
• ~7:40 a.m. ET — Codex lead Thibault Sottiaux posted: "We are aware that some Codex users are experiencing high error rates with 'model at capacity' and are working to bring things back to being stable" .
• 9:32 a.m. ET — A mitigation was applied and monitoring began .
• 9:48 a.m. ET — The status flipped to green: "All impacted services have now fully recovered" .

The outage affected Codex across its entire surface area, including the CLI, VS Code extension, and Desktop app . While the official incident clock showed about 3 hours of trouble, the practical disruption for users spanned a much longer window.

Developer Frustration Reaches a Boiling Point

For professional and Pro-tier subscribers, this wasn't a minor inconvenience—it was an active threat to their productivity. The reaction on the OpenAI Developer Community and X was visceral.

The core complaint wasn't just that the service was down, but how it failed. Users reported that Codex sessions would "exit mid-stream without saving state," forcing them to manually reconstruct lost context and redo work . A user in the UK summarized the sentiment vividly: "Making it impossible to work as you don’t know at what point Codex exits and it is going over and over the same thing finding what was done and what wasn’t. Totally unacceptable" .

The generic error message itself was a major source of anger. The advice to "try a different model" provided no actionable guidance when the primary model was unavailable and users had no way of knowing whether to retry, downgrade their reasoning effort, or simply wait .

Trust in OpenAI's communication was also damaged. Several users pointed out a gap between when the problems actually began—by community reports and personal experience—and when the official status page clock started ticking, a discrepancy that makes incident transparency feel unreliable .

Amid the frustration, a slice of developer gallows humor emerged. Influencer Matthew Berman created willcodexquotareset.com, a site that jokingly displayed a "94% chance of a Codex quota reset in the next 48 hours" . Digg’s sentiment analysis of the conversation around the event showed a split: 63.8% positive, with many thanking OpenAI for the quick fix, but a significant 36.2% negative, with users calling the service's reliability into question after a string of repeated outages .

A Troubled Year: Codex Reliability in 2026

The June 15-16 incident is not a one-off. It is the most visible spike in a year of recurring Codex degradation that began in earnest in early May 2026. A pattern of GPT-5.5 capacity saturation and rate-limit mismatches has surfaced repeatedly.

A timeline of major Codex events in 2026 shows a platform under persistent stress:

• April 23 — The "Selected model is at capacity" error is first widely reported on the Community forum, with users noting mid-session failures and no state preservation .
• May 7+ — A cluster of Codex latency and reconnect-loop degradations begins, with measurable worsening in public response times .
• May 12 — Following a "2x capacity" rollout, Codex App stream disconnects become common, ending with websocket closure errors .
• May 13 — GPT-5.5 experiences high error rates, disproportionately hitting free-tier users .
• May 14 — Codex Cloud and Code Review services suffer a high failure rate for 32 minutes .
• May 29 — A full-stack OpenAI outage brings down ChatGPT, the API, DALL-E, Codex, and Sora simultaneously .
• June 3 — Elevated error rates hit Codex, ChatGPT, and the Responses API for about 6.2 hours .
• June 11 — Elevated errors for GPT-5.5 in Codex appear in the evening, setting the stage for the main event .
• June 15-16 — The "Selected Model is at Capacity" error becomes a full-blown outage lasting roughly 3 hours .

The common thread is clear: GPT-5.5 demand keeps slamming into configured ceilings, whether from rate limits, reasoning-effort overload, or broader infrastructure strain. The June 16 fix was a rate-limit reset, a remedy that treated the symptom—the ceiling being hit—rather than the underlying mismatch between capacity and the model's popularity. Without a deeper infrastructure scaling solution, the error remains a likely returnee as more developers adopt Codex for high-intensity coding tasks.