On April 20, ChatGPT experienced a major global outage.
Just after 10 a.m. Eastern Time, failure reports on Down Detector began to surge. The peak in the UK exceeded 8,700 reports, and in the US it topped 1,900. In addition to the main ChatGPT interface, Codex, the API platform, and the Projects feature were also affected. Some users reported that ongoing work tasks were deleted due to the service interruption.
By 1 p.m., OpenAI had deployed a fix; services were largely restored by around 1:30 p.m. The total downtime lasted approximately three hours.
How OpenAI Responded
Throughout the incident, OpenAI's status page stated:
"We continue to investigate the issues with the listed services."
And:
"We have implemented mitigations and are monitoring the recovery."
There was no root cause explanation, no technical details, and no post-incident review.
This is not the first time. In several major ChatGPT outages in the past, OpenAI has not provided the kind of detailed post-mortem analysis that AWS or Google Cloud typically offer — reports that specify which component failed, why so many users were affected, and how it will be prevented next time. OpenAI has consistently provided the minimum information.
What Those 3 Hours Mean for Businesses
Three hours may not sound like a long time, but for teams that have embedded AI into their workflows, the impact is concrete:
- Ongoing Codex tasks were interrupted, losing progress
- Products integrated with the OpenAI API were completely non-functional
- Teams using ChatGPT for real-time customer requests were fully halted
- Some users' work content was deleted by the system due to the failure
Until businesses take the reliability of AI infrastructure seriously, the cost of such incidents will continue to be underestimated.
Why AI Services Are More Prone to Issues Than Traditional SaaS
Several structural reasons:
Highly concentrated traffic: ChatGPT is one of the most visited AI products globally, so any single point of failure has an extremely wide impact.
Complex inference chains: Large model inference is not like the simple request-response of traditional APIs. Load balancing and fault tolerance mechanisms are harder to design, and problems are more difficult to pinpoint quickly.
Fast iteration speed: Model updates and feature iterations are frequent, and each deployment can introduce new instability. AWS and Google Cloud have decades of reliability engineering experience; the engineering systems of AI service providers are still catching up.
The SLA Problem: The AI Industry Hasn't Seriously Addressed It Yet
Most mature SaaS providers offer a 99.9% SLA (Service Level Agreement), corresponding to no more than 8.7 hours of downtime per year; a stricter 99.99% SLA limits annual downtime to 52 minutes.
OpenAI currently has no public enterprise SLA commitment. Neither does Anthropic.
This means: a large number of businesses are betting their core workflows on a service with no clear reliability promise, and their contracts contain no compensation clauses for downtime. This is not a small issue, especially when AI has entered the critical path of production environments.
What to Consider If Your Business Depends on AI
Multi-vendor strategy: Connect critical operations to two sets of APIs, such as OpenAI and Anthropic. If one goes down, the other can still run. The implementation cost is not high, but routing logic must be prepared in advance.
Local model backup: Non-real-time tasks can use locally deployed open-source models as a fallback. For latency-tolerant scenarios, cache requests first and process them once the service recovers.
Contract terms: When signing enterprise contracts with AI vendors, include reliability requirements — even if the vendor currently has no standard SLA, there should be clear provisions for failure notification and compensation mechanisms.
Resilient code design: Timeout handling, retry logic, and fallback paths are not optional; they are basic engineering requirements when integrating external AI APIs.
This outage is not the first for OpenAI, and it will not be the last. As AI infrastructure matures, such incidents will continue to occur. The only question is: have the teams already using AI in production seriously designed contingency plans?
Sources: CocoLoop, ChatGPT was down — Live updates on the massive AI outage (Tom's Guide); ChatGPT was down for many — as OpenAI says it's 'monitoring the recovery' (TechRadar); OpenAI's ChatGPT Down for Thousands of Users (GV Wire); ChatGPT Down: Thousands of Users Hit by Major Outage on April 20 (TechGenyz)