OpenAI / GPT Relay Detection

Check for "leaked" fields in the response. OpenAI's official response only contains three token count fields. But if the relay secretly routes your request to a Claude backend and re-wraps it as OpenAI format, Anthropic's fields often leak through. When TokenAPI Scan catches these unexpected fields in the response, it marks them as a "critical issue" and drops the overall rating to yellow.

In short: genuine GPT responds according to OpenAI's own spec — counterfeit wrapper layers often leave traces from the other vendor.

Score is only one dimension. TokenAPI Scan also checks for "critical issues" — even if your score passes the threshold, as long as there is 1 critical issue (like the field leakage mentioned above), the overall conclusion drops to yellow "use with caution".

This usually means: the endpoint works and returns results, but it may not be running the model you think. If you're just chatting or writing everyday text, it might be fine. But for coding or serious tasks, we recommend switching providers.

Most likely the relay isn't broken — this key simply doesn't have GPT model access. Many multi-provider relays sell models grouped by vendor. Some customers buy Claude cards, others buy Gemini cards — if you haven't purchased GPT access, the models won't be available.

Before running detection, TokenAPI Scan quickly checks which models your key can use. If you're on the OpenAI page but see 0 GPT models, the system will automatically prompt: "This key can actually access X Claude models — want to test over there?" Just click through.

Because OpenAI doesn't include a "mathematical proof"-level anti-counterfeiting code in its responses. Claude has a cryptographic thinking signature that relays cannot forge, so TokenAPI Scan can give very definitive conclusions on Claude.

For OpenAI, we can only judge from indirect clues — "does the behavior look like GPT?", "are the field formats correct?", "are capabilities complete?" — so our stance is behavioral/protocol-level detection. We can tell you "looks normal" or "looks suspicious", but can't provide 100% mathematical proof. For definitive conclusions, we recommend testing the same relay with Claude model detection.