OpenAI API outage — what to do when the API is down

Your production application just started throwing 503 errors. Your monitoring dashboard is lighting up red. Before you panic-tweet or file an emergency support ticket, you need to confirm whether OpenAI's API is actually

OpenAI API outage — what to do when the API is down

Your production application just started throwing 503 errors. Your monitoring dashboard is lighting up red. Before you panic-tweet or file an emergency support ticket, you need to confirm whether OpenAI's API is actually down or if something broke on your end. Here's how to know for certain within 30 seconds, and what to do either way.

How to check if OpenAI is actually down

Start with the live OpenAI status monitor on this site — it updates every 60 seconds directly from OpenAI's official status feed and shows you exactly which services are affected. If you see "Operational" next to API services, the problem is likely on your end. If you see "Partial Outage" or "Major Outage", you're dealing with a platform-wide issue.

Cross-reference with status.openai.com for OpenAI's own incident reports. Look at the component-level breakdown — "API" might be green while "GPT-4 API" shows degraded performance. An outage affecting gpt-4-turbo doesn't mean gpt-3.5-turbo is down too. The status page breaks down Chat Completions API, Embeddings API, Assistants API, Fine-tuning API, and Batch API separately.

Check whether other developers are reporting issues. Search Twitter/X for "OpenAI API down" in the last hour. Scan the OpenAI Developer Community forums for new threads. If dozens of people are posting identical error codes in the past 15 minutes, you've confirmed a real outage.

Is it OpenAI or is it you

If the status page shows green but your requests are failing, isolate your setup systematically.

  • Test with curl directly — strip out your application layer completely. Run `curl https://api.openai.com/v1/chat/completions -H "Authorization: Bearer YOUR_API_KEY" -H "Content-Type: application/json" -d '{"model": "gpt-3.5-turbo", "messages": [{"role": "user", "content": "test"}]}'` from your terminal. If this works, your application code has the problem. If this fails, continue.
  • Check your API key — log into platform.openai.com, go to Settings → API keys. Verify your key hasn't been revoked. Check your usage limits under Settings → Limits. If you've hit your rate limit or billing quota, you'll get error 429 instead of 503. (See OpenAI API error 429 and other API errors for the full breakdown.)
  • Verify your network path — try from a different network entirely. If your production servers are behind a corporate firewall, test from your laptop on home WiFi. Some enterprise networks block api.openai.com at the firewall level without warning.
  • Check request formatting — a malformed JSON payload can trigger errors that look like outages. Validate your JSON structure. Confirm you're using the current API endpoint format (not deprecated v1 paths from 2023).
  • Review recent code changes — if this worked yesterday and broke today, what changed? New library version? Updated authentication flow? Environment variable typo?

What to do during an outage

When you've confirmed OpenAI's API is actually down, follow this protocol to minimise damage:

  • Implement exponential backoff immediately — if you're not already using it, add it now. Start with 1 second retry delay, then 2, 4, 8, up to 64 seconds maximum. Include jitter (randomisation) to prevent thundering herd when services recover.
  • Queue requests instead of dropping them — store failed requests in Redis, a database, or message queue. Process them when the API recovers. Don't return "Service Unavailable" to your users for every failed request during a 3-minute blip.
  • Switch to fallback responses — if you've built cached responses for common queries or a degraded-mode UI, activate it. Tell users "We're experiencing high demand. Your request is queued" rather than showing raw errors.
  • Monitor your retry budget — set a maximum retry count per request (typically 3-5 attempts). After that, fail gracefully and log for manual review. Infinite retries will hammer OpenAI's recovering infrastructure and delay everyone's recovery.
  • Don't spam support — OpenAI's support team can't restore API service faster by reading your ticket. Check the live OpenAI status monitor instead. Only contact support if the outage extends beyond the estimated recovery time on status.openai.com or if you're seeing issues that no one else reports.

What NOT to do: Don't regenerate API keys (wastes time and breaks your working services). Don't upgrade your plan thinking it'll bypass the outage (it won't). Don't delete and recreate your organisation (catastrophic if you have existing usage data).

Questions people actually ask

Q: Is OpenAI API down right now?

A: Check the live OpenAI status monitor on this page — it refreshes every 60 seconds from OpenAI's feed. If it shows "Operational" for API services, the API is up. If you're still having issues, see ChatGPT not working — troubleshooting for diagnostic steps.

Q: How long do OpenAI API outages typically last?

A: Most partial outages resolve within 15-45 minutes. Major incidents can extend 2-4 hours. OpenAI posts estimated recovery times on status.openai.com during active incidents.

Q: Will OpenAI refund API credits for downtime?

A: OpenAI doesn't automatically issue credits for outages. Their SLA (if you have an enterprise agreement) may include uptime guarantees, but standard pay-as-you-go accounts don't have refund provisions for brief outages.

Q: Should I switch to a different model during an outage?

A: Yes, if GPT-4 API is down but GPT-3.5 Turbo works, switch temporarily. Check the component-level status to see which specific models are affected.

What to remember

  • The live status monitor on this site updates every 60 seconds — bookmark it for instant outage confirmation
  • Test with raw curl commands to eliminate your application as the problem source
  • Implement exponential backoff with jitter before the next outage, not during it
  • Queue failed requests rather than dropping them — users tolerate delays better than errors
  • Never regenerate API keys or upgrade plans during an outage — you're solving the wrong problem
  • Check status.openai.com for OpenAI's official incident timeline and recovery estimates

---

*openai-support.com is an independent resource, not affiliated with OpenAI Inc.*

Related help