OpenAI / GPT relay check

Clodex checks whether a relay behaves like a real OpenAI Chat Completions endpoint and looks for fingerprints of Claude / Gemini backend rewrapping.

Behavior / protocol level: OpenAI responses do not expose a Claude-like cryptographic signature, so this is not a mathematical proof of model identity.

Relay base URL

OpenAI-compatible root URL, e.g. https://api.example.com or https://api.example.com/v1.

API key

Used only for this check; not written to reports or persisted.

Target model

Check depth

Long-context check Extra 32k -> 100k -> 200k probes. This can cost more and is paid by your API key.

Extreme context Tests closer to the advertised model limit. Can be expensive for large models.

Chat Completions

Checks ids, object types, choices, finish_reason and usage shape.

Capability integrity

Runs tool call and JSON schema probes to catch simplified adapters.

Usage fingerprints

Flags suspicious Claude / Gemini fields inside OpenAI-shaped responses.

FAQ

Can this prove a GPT model is genuine?

Not cryptographically. It verifies protocol behavior and catches common backend-conversion fingerprints.

Why can a 75 score still look risky?

Critical issues can downgrade the verdict even when the numeric score is passable.

More OpenAI relay questions: /faq#openai