By Byron V. Acohido

Millions of people rely on ChatGPT Work. For more than five hours Monday, it would not run. I was one of the people watching it fail.

The trouble began at 8:04 a.m. Pacific and ran through the heart of the American workday. OpenAI declared recovery at 1:28 p.m. Its incident record lists elevated latency, elevated errors, a mitigation and a recovery. It names no cause, no scope and no count of who was affected.

So Tuesday morning I put the questions to the company’s own public-facing tool.

What surfaced was not in the incident record. Monday was also the day OpenAI retired GPT-5.4 and GPT-5.4 Mini from Codex and other work surfaces authenticated through a ChatGPT account. The company announced that change a month ago and gave customers ample notice. It did not mention the change once while its paid work product was failing.

And when I returned to a restored ChatGPT Work, my session opened on an engine I had never seen before, at a reasoning level I never use.

OpenAI has disclosed no connection between the model transition and the outage. It has disclosed no connection because it has disclosed nothing at all.

What OpenAI said

The company’s status page moved in three short steps.

Engineers first flagged elevated latency in ChatGPT Work. Roughly 17 minutes later the notice was upgraded to elevated errors, with a callout that Plus subscribers were hit hardest and that Work mode was unavailable to them. A later update widened the scope to users across multiple subscription plans who could not start or continue tasks.

Then came a mitigation, then recovery. Total elapsed time was about five and a half hours.

Alongside the status page, the company posted a single line to X in mid-outage: “ChatGPT Work isn’t working right now. Sorry about that, we’re working on a fix!”

That is the entire public account. OpenAI has not said what failed, whether the trigger sat inside ChatGPT Work or in shared infrastructure, whether unfinished work or task state was at risk, or whether it intends to publish a postmortem.

Nor has the company answered the press. Folha de S.Paulo, which covered the outage as it unfolded, reported that OpenAI did not respond to its inquiry before publication.

Reticence within a day of a complex outage is defensible. Root-cause investigations take time. Premature explanations can be wrong, can expose internal architecture and can seed narratives that are hard to correct later.

The reticence still carries a cost for a company asking consumers, professionals and enterprises to wire its engines into daily operations.

An unknown blast radius

The number of people affected cannot be calculated from what is public.

Downdetector logged more than 2,000 reports by 9:01 a.m. Pacific, and the count climbed past 10,000 as the morning went on. Complaints came in from the United States and Britain, among other places.

Brazil ran on the same clock. Downdetector complaints there began climbing at 12:03 p.m. Brasilia time, which is 8:03 a.m. Pacific, and peaked at 2:33 p.m. with 423 connection reports. Brazilian outlets tracked the disruption through the afternoon, and Folha noted signs of trouble in other countries.

Those are voluntary complaints. They set a floor, not a headcount. Most people who lose a service never file a report, and one frustrated user can file several.

The potential exposure is large. OpenAI said on July 31 that its models now reach more than 1 billion active users and more than 2 million businesses. That figure spans all of the company’s interfaces, including ChatGPT, Codex and ChatGPT Work. The last paid-tier count the company disclosed was 50 million subscribers in February.

Monday’s failure did not touch all of that. It touched paid work surfaces, and the status page described the affected group only as “multiple subscription plans.” No geographic breakdown. No plan-level breakdown. No estimate of accounts touched.

The missing denominator matters. OpenAI can determine which accounts, regions and task types threw errors. Its customers cannot.

A same-day switch

Monday was also a scheduled deployment date.

OpenAI retired GPT-5.4 and GPT-5.4 Mini for anyone signed in with a ChatGPT account, including Codex sessions authenticated that way. Both models remain available through the API and through Codex sessions authenticated with an API key. The company announced the retirement on July 31 and told users to swap in GPT-5.6 Terra for GPT-5.4 and GPT-5.6 Luna for GPT-5.4 Mini across saved settings, managed configurations, custom agents and scheduled tasks.

A model retirement and a five-hour outage landing on the same calendar day may be coincidence. Large platforms retire models routinely, and OpenAI has not linked the two events.

It has not unlinked them either. The incident record makes no reference to the transition at all, which leaves customers to notice the overlap on their own.

A dial I never set

When service came back, I went looking at the machinery.

My next ChatGPT Work conversation opened on GPT-5.6 Terra Light. Two things stopped me.

I had never seen the word Terra in my engine selector. And I do not run Light. I work at medium and high reasoning levels, and I have for months.

Terra is also the specific engine OpenAI named as the replacement for the model it retired that morning. I found the setting only because I opened the engine-selection control mid-conversation. I preserved screenshots.

OpenAI’s release notes describe Sol Medium as the normal default. The live control in my account offered five graduated presets: Terra Light, Sol Light, Sol Medium, Sol High and Sol Extra High.

This is one account and one observation. It does not establish that the outage, the retirement and the changed setting share a cause. Lost preference state, a migration rule, a staged interface rollout, a capacity-management measure or an ordinary software defect could each produce the same result.

It does raise a question that OpenAI can answer and I cannot: did recovery from the outage, or the retirement of GPT-5.4, change model or reasoning settings for any users, and were those users told?

The question is not academic for people who use ChatGPT to research, analyze records, edit copy or produce client deliverables. Different engines at different reasoning levels produce materially different work. A customer who selects a given engine has reason to expect that the choice persists, or that a change arrives with notice.

That expectation is harder to hold when the control itself keeps moving. The graduated engine bar has been reworked more than once during 2026, and OpenAI’s public changelog does not preserve a version-by-version record of what the bar displayed and when. Reconstructing my own interface history is not possible from the company’s documentation.

Twice in four days

Monday was not an isolated event.

The same product logged an incident on Aug. 27, described as elevated error rates in Workspace Agents and ChatGPT Work on web and mobile. Two disruptions to the same paid work surface inside four days is worth noting on its own.

The broader cadence is heavier. StatusGator counted 61 ChatGPT incidents over the last 90 days at a median duration of about two hours and 40 minutes, and 349 OpenAI incidents since January 2025. Most are brief. IsDown puts the median resolution time at two hours and 28 minutes across 208 incidents logged since October 2025.

Monday ran roughly twice that median. Neither Monday nor Aug. 27 has a published cause.

Why the silence

OpenAI is running a familiar incident-response calculation.

Say too much too early and the company risks publishing a wrong diagnosis, exposing exploitable detail or letting an incomplete account harden into accepted fact. If a security issue is involved, a responsible investigation may require coordination with vendors, customers or law enforcement before anything goes out.

Say too little and a different risk opens. Speculation fills the space. Customers lose information they need for continuity planning. Restoring convenience starts to look like a substitute for accountability.

OpenAI has taken the minimal route. Its updates said engineers identified elevated latency and errors, were implementing a mitigation and had restored service. The company has not committed to an after-action report. It has not stated whether customer data and task state came through intact.

None of that proves concealment. It falls short of what enterprises now expect from cloud, communications and other operationally important providers.

An infrastructure-grade response would eventually name the root-cause category, the scale and duration of customer impact, whether data or unfinished work was at risk, what corrective control was added and whether any automatic fallback or configuration change occurred. Before the investigation closes, the company could still say what it knows, what it does not know and when customers should expect the next report.

From tool to utility

The larger issue is dependency.

ChatGPT is no longer a chatbot people visit for occasional answers. OpenAI markets ChatGPT Work as a system that gathers organizational context, executes multistep assignments and produces finished documents, presentations, spreadsheets, software and research.

That concentrates work, context and decision support inside a platform customers cannot repair, audit or readily replace when it stops.

A single outage builds no regulatory case. Repeated outages paired with thin disclosure eventually will.

Compliance teams and regulators will get around to asking whether foundational AI providers should carry clearer incident-reporting standards, auditable model provenance, continuity controls and service commitments. They will also ask whether customers deserve notice when an AI work platform changes the engine handling their assignments.

OpenAI is racing formidable competitors to ship capabilities and hold customer loyalty. New features are visible and commercially rewarding. Redundancy, graceful degradation, rollback discipline and transparent incident review stay invisible until they go missing.

Monday’s failure poses a bigger test than whether ChatGPT Work could be brought back online.

The test is whether foundational AI companies can grow into accountable infrastructure providers while keeping the speed that made them competitive.

OpenAI has restored the service. It has not yet supplied the explanation.

Pulitzer Prize-winning business journalist Byron V. Acohido is dedicated to fostering public awareness about how to make the Internet as private and secure as it ought to be.

(Editor’s note: the questions in this account were put to ChatGPT itself, the channel OpenAI makes universally available. Its answers are not official company statements, and the figures reported here were verified independently against OpenAI’s status record, release notes and public announcements, against outage trackers, and against contemporaneous news coverage. OpenAI did not respond to Folha de S.Paulo’s request for comment and has published nothing about Monday’s failure beyond its status record and a one-line post to X.)

September 1st, 2026 | My Take | Top Stories