OpenAI's April 20, 2026 Disruption: What It Means for Apps
What OpenAI's own write-up says about the April 20, 2026 disruption, what it leaves open for API callers, and four lessons for your app.
OpenAI published a write-up of a service disruption on April 20, 2026. This post covers what it says, what it leaves open for API callers, and what to take from it. Read the write-up itself before you rely on any summary.
What OpenAI says happened
On OpenAI's status page the incident is titled "Users unable to load ChatGPT, Codex and API Platform" and marked as a partial outage. The write-up says the disruption ran from approximately 7:18 AM to 9:00 AM PDT and affected ChatGPT availability for some users. That window is 14:18 to 16:00 UTC.
Some users got errors in ChatGPT. Some were asked to log in again or hit login and signup failures. Some had login or signup errors in other products, including the Codex app and platform.openai.com. The heaviest impact was in South America, India and Europe.
According to the write-up, the cause was an overload in an internal authentication edge dependency that authorizes ChatGPT requests. A temporary backend disruption exposed insufficient headroom in a critical authentication path, and retry, caching and routing behavior amplified the overload. Authentication systems had recovered by about 8:20 AM PDT, ChatGPT edge availability by about 8:40 AM PDT, and overall ChatGPT metrics by about 9:00 AM PDT.
Paying full price on every request?
PromptUnit checks each request your app sends and moves it to a cheaper model when one is just as good for that request. You change two settings, and your own provider key stays. Free until it has saved you money, then 20% of what it saves.
Try it free →What it means for an app
The write-up is about ChatGPT and sign-in flows. It does not say whether direct API requests, the kind your server sends with an API key, failed. The incident title also names "API Platform", but the write-up gives no API error rates or symptoms.
- ChatGPT and Codex app users, including your own team, are in the impact list.
- Anyone who needed to sign in to platform.openai.com during the window may have had trouble.
- For server-side API calls, the write-up gives no answer. Check your own logs for that window, with margin on both sides.
Lessons for your app
- Know which surface you depend on. ChatGPT, the Codex app, platform.openai.com and the API are different surfaces. Your logs, not a headline, tell you which one hurt you.
- Log every provider error with a timestamp. Keep the status code, error code and time, so you can match it to a provider's write-up later.
- Cap and space out retries. The write-up lists retry behavior among the things that amplified the overload. A client can add load to a struggling provider the same way. Limit attempts and add a random delay. Honor
Retry-Afterwhen the API sends it, as OpenAI's error code docs say to. - Show users a clear state. OpenAI's Python SDK waits up to 10 minutes by default (README). Set a shorter timeout so a visible error or a degraded mode replaces a spinner.
Where to go next
For the how-to, with an error table, timeouts, sample code and tests, read LLM Provider Failover Guide: Errors, Timeouts and Testing.
Where PromptUnit fits
PromptUnit is a proxy you point your AI API base URL at. If a cheaper model it chose fails, it retries on the model you asked for. It does not switch to another provider when your provider is down, so failover between providers is still yours to build. It is free until it has saved you money, then 20% of savings.