OpenAI’s API surface has grown a lot since the original GPT-3.5 days. There is now the Responses API, the Vector Store API, the Images API for gpt-image-2, and the older Chat Completions still around. Different endpoints, different error styles. This is the 2026 short list of what actually breaks, and the fix for each. 😺

These errors are not AI Engine bugs. They come back from OpenAI when something is off with your account, your network, or your request shape. AI Engine surfaces them as-is so you can act on them.

Diagram of the OpenAI cloud with three coral arrows pointing to Billing, Network, and Deprecated Model buckets.

insufficient_quota / You exceeded your current quota

This is the most common one and it is almost always billing, not rate-limiting. OpenAI moved to prepaid billing in 2024: you keep a positive balance, the API works. The balance hits zero, the API stops.

  • Check your balance in the OpenAI billing dashboard.
  • Add credits (minimum $5).
  • Set both soft and hard usage limits so you do not get surprised next month.
  • Wait 5 to 10 minutes after topping up before testing.

If you have credits and still see insufficient_quota, regenerate your API key. OpenAI sometimes invalidates the cached auth state after billing changes.

Rate limit error (429)

You are sending requests faster than your tier allows. The error body usually includes the current rate, your limit, and the model. If you see something like Limit 10000.000000 / min, that is your monthly tier ceiling.

  • For occasional spikes: AI Engine retries with exponential backoff automatically. You should not see the user-facing error in normal use.
  • For chronic 429s: you have outgrown your tier. Spend a bit on the API for a few weeks and OpenAI will bump you up automatically.
  • For image generation: gpt-image-2 has its own separate rate limits, lower than text. Plan accordingly if you batch-generate.

Invalid API key / 401 Unauthorized

Your API key is wrong, expired, or has been revoked. The fix is mechanical:

  • Open AI Engine → Settings → API Keys. Make sure there are no stray spaces or quotes around the key.
  • If that does not help, generate a fresh key in the OpenAI API keys page and paste it in.
  • If you have project-scoped keys, make sure the project actually has access to the models you are calling.

No tool output found (Responses API)

Specific to the Responses API + Vector Store flow. The model called a tool, the tool returned nothing, and the API errors out. Common when your vector store is empty, embeddings are stale, or the file you uploaded was unreadable.

We wrote a dedicated piece on this one: Fix: No Tool Output Found. Five-minute read.

Model not found / Invalid model

You are asking for a model that does not exist for your account, has been deprecated, or is misspelled. As of mid-2026, OpenAI has quietly retired most pre-GPT-4 models. Anything with davinci, curie, babbage, ada in the name is gone.

  • For text: gpt-4o, gpt-4o-mini, gpt-4.1, o3, o3-mini are the current workhorses.
  • For images: gpt-image-1, gpt-image-1.5, gpt-image-2. The dall-e family is deprecated for new accounts.
  • For embeddings: text-embedding-3-large or text-embedding-3-small.

Check the current OpenAI models documentation if something feels off.

Context length exceeded

Your prompt plus the model’s response would not fit in the context window. Less of an issue in 2026 than it used to be (modern models have 128K to 1M token windows), but it still happens with huge embedded contexts or long chat histories.

  • Reduce the chunk size or count when AI Engine embeds context.
  • Trim the conversation history sent with each chatbot request (AI Engine has a setting for this).
  • Move to a larger-window model if you genuinely need it.

503 / Model is currently overloaded

OpenAI’s servers are saturated for that model. Usually a 30-second to 5-minute issue. AI Engine retries automatically, so you mostly will not notice. If you do, switch to a sibling model for a few minutes (e.g. gpt-4o-mini while gpt-4o is overloaded) or just wait.

cURL error 28 / Connection timeout

The request never reached OpenAI. This is your network, not theirs. The three common causes:

  • Geo-blocked. OpenAI does not serve some regions (notably mainland China). You will need a proxy or an OpenAI-compatible regional gateway.
  • Security plugin. Wordfence or similar blocks outbound HTTPS by default on some configs. Whitelist api.openai.com. We have a piece on this.
  • Host firewall. Some shared hosts block outbound connections to anything that is not their own services. Open a support ticket or move to a real host (our hosting recommendations).

In short

Most OpenAI errors are billing, network, or a deprecated model. Check those three first and you will fix 90% of them in under a minute. The Responses API has its own quirks (vector stores, tool calls); when those break, the dedicated guide is the place to look.