When AI Engine shows an error, it almost always comes from the AI provider: a balance at zero, a key that doesn’t match, a rate limit, a retired model. Here are the 8 errors people actually hit today with OpenAI, Anthropic (Claude), Google (Gemini) and OpenRouter, and the fix for each. 😺
In 30 seconds
- Test while logged in as an admin: visitors only see a generic “Oops! Something went wrong”, admins see the provider’s real message.
- Most errors are billing: “no credits remaining” (OpenAI) or “credit balance is too low” (Anthropic).
- Keys live in Settings → AI → Environments for AI, with a Quick Test button.
- “Usage limit for now” is AI Engine’s own cost guard, not the provider: see Insights → Limits.
- Timeouts and blocked requests come from your server, not from the AI.
First: see the real error
Visitors who aren’t admins only ever see: “Oops! Something went wrong on the server. Please try again, and if you are the site developer, check the PHP Error Logs for details.” That’s on purpose, so they never see your account details. To see what’s really wrong, try the chatbot while logged in as an admin: the message then starts with the provider’s name, like OpenAI (Responses API): …, From Anthropic: …, From Google: … or OpenRouter: …. For more detail, turn on Settings → Dev Tools → Queries Debug, or read your PHP error logs.
1. “You have no credits remaining” (OpenAI)
OpenAI is prepaid: the account behind your key has no balance, or the key belongs to another organization or project. Add credits at platform.openai.com. If your balance is already positive, create a new key in the same organization and project, and paste it in Settings → AI → Environments for AI. Note that Quick Test won’t catch this one: it only lists the models, which needs no credits.
2. “Your credit balance is too low” (Anthropic)
Same thing on Claude’s side: the message reads From Anthropic: Your credit balance is too low…. Buy credits in the Claude Console. A Claude.ai subscription doesn’t count: the API is billed separately.
3. Invalid or missing API key (401)
On Anthropic, Quick Test says “Connection failed: Invalid API key”. On OpenRouter: “The API key was refused by OpenRouter (HTTP 401)”. OpenAI passes its own message through. Paste the key again in Settings → AI → Environments for AI (no spaces around it) and press Quick Test. If the key won’t even save, it’s usually ModSecurity on your host blocking the settings request: ask your host to whitelist it.
4. Rate limit (429)
Something like “tokens per minute… Requested 122,620”. Your provider tier is low and the prompts are big, usually because of long chat history or a lot of Knowledge context. AI Engine doesn’t retry on its own, so lower the load: reduce Max Messages and Context Max Length in the chatbot, lower Max Results in Knowledge, or move to a higher tier with your provider (they raise it as you spend).
5. Model not found or retired (404)
“The API returned 404 for … This usually means the model is not valid for this provider.” On Gemini, AI Engine adds that the model may have been retired by Google. Pick a current model in the chatbot, or in Settings → AI → Default Environments for AI. Models do retire on schedule: the GPT-5 line on December 10, 2026, Claude Sonnet 4.5 on November 30, 2026. On OpenRouter, “No endpoints found” means the same: that model was dropped, choose another.
6. “You have reached the usage limit for now”
That’s not the provider: it’s AI Engine protecting your bill. By default, each visitor who isn’t logged in gets 3 questions a day. Change it in Insights → Limits, per user type.
7. Timeouts and blocked requests
“Your server replied with an HTTP … error instead of an AI reply. The request probably timed out or PHP crashed…” The AI never got a chance: PHP’s max_execution_time, a proxy limit, or a firewall (Cloudflare, ModSecurity, some hosts like Hostinger) stopped the request. Raise the PHP limits (AI Engine itself waits up to 300 seconds, adjustable with the MWAI_TIMEOUT constant) and whitelist /wp-json/mwai* in your firewall. If your security plugin blocks outgoing requests, allow the provider’s API too (how to whitelist external requests).
8. “Your session has expired. Please refresh the page”
WordPress security tokens (nonces) last about a day. On a page served from cache, the chatbot can get one that’s too old. Refresh the page, and exclude the pages with a chatbot from your page cache. Same family of problem as “Cookie check failed”.
“Overloaded”: just wait
Anthropic’s 529 Overloaded and Gemini’s 503 UNAVAILABLE are on their side and usually last minutes. Retry later, or switch the chatbot to another model for a while. Is ChatGPT, Claude or Gemini down? tracks the incidents.
Errors you won’t see anymore
If you’re reading an old forum thread: “text-davinci has been deprecated”, “That model is currently overloaded” from 2023, “You exceeded your current quota” on a free trial (OpenAI is prepaid now), max_tokens or temperature errors on new models (AI Engine adapts the parameters by itself), and Assistants errors (OpenAI shut the Assistants API down in August 2026). For the Responses API and vector stores, see Fix: No Tool Output Found.
Still stuck? The common issues ↗ in the docs go further, and our support is here for the rest.