Image: Unite.AI / unite.ai OpenAI's ChatGPT, Codex, and API Have Gone Down Four Times in Four Days
OpenAI's services went down again Saturday morning, marking the fourth disruption in four days across ChatGPT, Codex, and the developer API. The company resolved it within an hour, but the pattern of roughly 18 incidents per month is drawing attention as more developers run autonomous agents that can't easily pause mid-task.
OpenAI’s status page lit up again Saturday morning. Around 10:54 AM UTC on July 25, the company acknowledged elevated error rates across ChatGPT, its developer API, and Codex simultaneously, marking the fourth service disruption in four days.
Users got 503 errors labeled biscuit_baker_service_me_circuit_open, which blocked requests from reaching OpenAI’s servers. The impact was broad: 15 components on the ChatGPT side, 12 across the API, and 4 on Codex. OpenAI moved from investigating to monitoring within about an hour, saying it had applied a mitigation and was watching for recovery.
The pattern
This wasn’t isolated. The four-day run included elevated API errors and Codex Review issues on Thursday, ChatGPT and API image generation problems on Wednesday, and a separate major outage on July 23, two days earlier. There was also a significant outage in April.
Zooming out, OpenAI’s status history logs roughly 166 incidents over nine months since autumn 2025. That’s about 18 disruptions per month across a platform serving 900 million weekly active users and 10 million users running agentic workflows.
Why it matters more than it used to
A chatbot going offline for an hour is annoying. An autonomous agent going offline mid-task is a different problem. Agents coordinating deployments, running test suites, or managing code review queues don’t have a natural pause point. When the service drops, their work either rolls back, stalls, or leaves things in a partially completed state that someone has to clean up manually.
As more developers move from “AI-assisted editing” to “AI running pipelines,” reliability becomes a harder requirement. Mitigation is possible with retries and fallbacks, but it adds engineering overhead, and it only helps if the downtime is short enough that retried operations stay coherent.
OpenAI hasn’t commented publicly on the frequency of recent incidents or what’s driving them at this scale.
Sources: The Next Web · Unite.AI · BleepingComputer