FAQ
Common developer questions. For billing, SLAs, and contract questions, see the .
Which OpenAI SDK version do you support?
Any version from v1.x onward. The request and response shapes match the Chat Completions spec — older versions may not support newer parameters like tool_choice or response_format, but base functionality works across versions.
Are my prompts stored?
No. Prompts and responses pass through in real time and are not logged. We only record metadata: model used, token counts, cost, and timestamp for billing and abuse detection.
Do you support automatic fallback between providers?
Not yet. If a specific model is unavailable, the API returns a clear error and you can retry against another model. Automatic fallback routing is on our Q3 2026 roadmap.
What happens if I run out of credits mid-request?
The in-flight request completes and is billed from any remaining balance. Subsequent requests return 402 until you top up. Spending caps per key can prevent this from happening on specific keys.
Is there a sandbox or test mode?
Not a separate sandbox. Use the Playground in the dashboard to test models interactively, or create a key with a low spend cap (e.g. $1) as a safe dev key.
I asked the model "what model are you?" and it gave the wrong answer. Is the routing broken?
Almost certainly not. Models don't have a reliable lookup for their own identity — they generate the most statistically likely answer given their training data, and the most likely answer to "what model are you?" is whatever appeared most often in the training corpus. For newer models, that's usually an older sibling ("DeepSeek V3", "GPT-4") or a generic "ChatGPT/OpenAI" pattern that flooded the internet for years. It's training-data contamination, not a routing bug. To verify the actual model that served your request, look at the emerald "served by:" chip below the response in the Playground — that value is set by the gateway from the upstream response metadata and is authoritative. If your application needs the model to identify itself reliably, add an explicit instruction in your system prompt ("You are DeepSeek V4 Pro made by DeepSeek."). We deliberately do not inject this by default — it would mask the rare case of actual mis-routing.