API FAQ

High-frequency questions about API calls. Match your symptom, then follow the links for details.

Getting 401 Unauthorized?

• Wrong key: check you are using the full key saved at creation (the masked list entry is not the full key);
• Key deleted or expired: confirm its status on the "API Keys" page and recreate if needed;
• Malformed header: the OpenAI-compatible protocol expects Authorization: Bearer <key> — do not drop the Bearer prefix;
• More in Authentication.

"Model not found"?

• Misspelled model ID — copy the exact ID from the model marketplace instead of typing it;
• The model does not exist on the platform or is currently unavailable — check the card exists and is available in the marketplace.

Timeouts or slow responses?

• Long-context / long-output requests are inherently slower; check whether max_tokens is set unnecessarily high;
• Set sane client timeouts and retries with exponential backoff;
• For interactive UIs use "stream": true — first bytes arrive sooner and perceived latency drops.

Frequent 429s?

You are hitting rate limits. Lower concurrency, add backoff intervals; for sustained high load, contact the platform about quotas. See Error codes.

Empty or truncated replies?

• finish_reason is length: the reply hit the max_tokens cap — raise it;
• Check that no messages entry carries an empty content;
• Some models constrain input formats (e.g. image modality requirements) — see the model card notes.

How do I confirm a call actually succeeded?

Two checkpoints: HTTP 200 with choices / usage in the response; and the matching record (with token usage and cost) on the console logs page. When both agree, the pipeline is healthy.