Thousands of companies have built their products on OpenAI's API. They've staked their businesses, their investors' money, and their customers' trust on a platform that has proven catastrophically unreliable. The pattern is clear: outages during critical moments, silent model changes that break applications, rate limits that appear without warning, and a support system that treats paying customers like an afterthought.
This page documents the ongoing API reliability crisisânot as isolated incidents, but as a systemic pattern that makes OpenAI's platform a dangerous foundation for any serious business.
A Y Combinator-backed startup built their entire product on OpenAI's API. They raised $2.3 million and signed contracts with enterprise clients. Then the Black Friday outage happened.
"We promised 99.9% uptime to our clients. We based that on OpenAI's SLA. On Black Friday, their API was down for six hours. Our clients' customer service went dark during their biggest sales day. We lost three enterprise contracts that week. By December, we couldn't make payroll. We're shutting down in January."
The founders are now advising other startups: "Never build your core product on a single API provider, especially not OpenAI."
A legal technology company used GPT-4 to analyze contracts and extract key terms. Their system had been tested extensively and was delivering accurate results. Then OpenAI silently updated the model.
"Without any notice, the model's output format changed. Our parsers broke. But worseâthe model started hallucinating contract terms that didn't exist. A client nearly signed a deal based on AI-generated fiction that we didn't catch in time. We had to shut down the product and do a complete audit. OpenAI's response? 'Models are continuously improved.' No changelog. No warning. No apology."
The company is now building their own models in-house, despite the cost, because they can't trust OpenAI's API stability.
An educational platform served 50,000 students using ChatGPT for tutoring. They'd carefully calculated their API costs and rate limits. Then OpenAI changed the rules.
"We were well under our rate limits. Then one day, 429 errors everywhere. We contacted supportâthey said our 'usage pattern' triggered automatic throttling. What pattern? They wouldn't tell us. We had 50,000 students in the middle of exam prep who suddenly couldn't use our platform. It took two weeks to resolve. Two weeks of students failing exams because OpenAI's rate limiting is a black box."
The platform now maintains fallback systems with three different AI providers, tripling their infrastructure costs.
A marketing agency woke up to a $47,000 API bill for a month where they'd budgeted $4,700. OpenAI's billing system had malfunctioned.
"Our usage didn't change. Our code didn't change. But OpenAI billed us 10x our normal amount. When we disputed it, they took three weeks to respond. During that time, they'd already charged our card. Getting a refund took two more months. We nearly went bankrupt because of their billing bug. And they offered no compensation for the stress, the bounced payroll, or the damage to our credit."
A healthcare platform used ChatGPT to help patients schedule appointments and answer medical triage questions. During an 8-hour API outage, patients couldn't access care.
"We had a patient call our backup line in tears. She'd been trying to schedule an urgent appointment through our AI system all day. By the time she reached a human, the specialist she needed had left for the day. She had to go to the ER instead. That ER visit cost her $3,000 and hours of waitingâall because OpenAI's API went down and we trusted it for critical healthcare functions."
The platform has since implemented mandatory human fallbacks for all patient-facing functions, doubling their operational costs.
Three startups at a major accelerator demo day had their presentations ruined when OpenAI's API went down during their live demos.
"I was on stage in front of 200 investors, about to show our AI product in action. The API returned an error. Then another. Then timeout. I had to apologize to a room full of VCs and explain that our productâwhich had worked perfectly for monthsâcouldn't demo because OpenAI was having issues. We didn't get funded. Two years of work, dead on stage because of API reliability."
The accelerator now advises all startups to have offline demo modes that don't depend on live API calls.
Promised SLA: 99.9% uptime
Actual measured uptime in 2025: 99.2%
That 0.7% difference = 61 hours of downtime per year
Average Response Time to Support Tickets:
Free tier: 14 days
Pro tier: 7 days
Enterprise tier: 3 days
Critical outage during business hours: Still 3 days
Model Changes Without Notice in 2025: 6 major changes that broke production systems
Rate Limit Disputes Resolution Time: Average 23 days
Billing Error Refund Time: Average 67 days