ChatGPT Image Generation: Six Days on One Open Incident and No Way to Predict a Failure
OpenAI has kept one image generation incident open in ChatGPT since 21 July 2026. The matching incident on the developer interface closed the same evening.
The Bright Recap
OpenAI opened an incident titled "Image generation unavailable in ChatGPT" on 21 July 2026 and it was still listed as investigating on 27 July, six days later. The entry reports elevated errors and a mitigation being implemented, which describes intermittent failure rather than a service that has stopped.
A separate incident covering image generation through OpenAI's developer interface was raised and fully resolved on the evening of 21 July. A further incident affecting ChatGPT conversations reached the monitoring stage within about a day. Only the entry attached to the consumer application has stayed open.
To know more about this topic, read our related articles:
- The risk of depending on too few
- What happens when a request is flagged
- When the manual process comes back
- What OpenAI charges and what it promises
- Financial technology explained
Bright Answers
Why does ChatGPT image generation work sometimes and fail others?
OpenAI's open incident describes elevated errors rather than a total loss of service, so a share of requests completes normally while others return an error. The same account can succeed in the morning and fail in the afternoon, which is why individual reports of the feature working do not mean the incident has been resolved.
How do I check whether the problem is mine or OpenAI's?
OpenAI publishes live incidents by component at status.openai.com, and each entry names which part of the product is affected. If image generation appears there, rewriting the prompt, clearing the cache or logging out will not change the outcome.
An error you can reproduce is a problem you can plan around. OpenAI has spent six days on the other kind, keeping an incident titled "Image generation unavailable in ChatGPT" open since 21 July with the same sentence attached to it about elevated errors and a mitigation being implemented. Elevated errors means intermittent, so the feature has been producing images for some people while returning failures to others throughout the week. The version of the same fault on OpenAI's developer interface was raised, mitigated and closed on the evening it appeared.
The asymmetry is visible on a public page
The developer-side entry ran through investigating, monitoring and resolved within hours, with OpenAI reporting that error rates had returned to normal levels. A further incident covering file uploads and image generation on 22 July also closed. The entry attached to the consumer application has not moved past investigating in six days. Paying developers have contractual expectations and a rival supplier one configuration change away, which is a difference we have looked at before in what OpenAI charges alongside what it commits to.
Intermittent is the expensive kind for a business
A subscriber who loses an afternoon to a failed image has lost an afternoon. A company that has built image generation into a publishing calendar, a product listing workflow or a client deck loses something harder to price, because intermittent failure defeats scheduling. Work gets planned on the assumption the tool functions, since it functioned yesterday, and the gap only surfaces when an asset is missing at the point of delivery. Businesses recognise this pattern from any automated step that quietly hands the job back to a person, and degradation rather than absence is now the ordinary shape of an AI failure, as it is when a request is flagged and routed to a weaker model.
Three things cost nothing and remove most of the pain. Read the component-level status page before touching the prompt, because no rewrite fixes an incident and retrying during one buys delay. Build the deadline around a buffer rather than around a first attempt succeeding. Keep a second image tool that a colleague can already operate, which is the same discipline any financial technology process needs when it runs on one supplier, and the same exposure that appears whether the single dependency is a network or a model.
OpenAI closed the developer version of this fault the evening it opened. The version affecting people using the app has been carrying the same sentence for six days, and the only reliable way to know whether it will work is to ask it.
Editor's note
Every piece goes through careful verification, but mistakes can happen. Readers who spot an error or have additional information can write to rosalia@thebrightminded.com.