Auditing GPT-5.6 APIs for Model Mixing and Shadow Routing

What it does

gpt56_api_detector is a locally hosted Python tool that stress-tests an API endpoint claiming to serve specific GPT-5.6 variants—Sol, Terra, or Luna. It fires fixed behavioral probes and checks so-called “Juice” budget fingerprints to see if the answers match the declared model’s known distributions, or if they look like a different variant, a rewritten output, or a non-GPT backend entirely. Everything stays on 127.0.0.1; the tool never writes your API key to disk or reports.

The interesting bit

The project treats model verification like a forensic accounting audit. It watches for oddly specific tells—such as the numbers 32 or 48 being crudely rewritten to 40—and compares how the endpoint behaves under native Codex shapes versus plain chat contexts. The README is admirably upfront that a failing grade does not automatically mean your reseller is cheating; it could just be upstream risk controls, rate limits, or temporary routing layers acting up.

Key highlights

• Runs entirely local; API keys live only in page memory and backend RAM until shutdown.
• Uses two independent evidence tracks: deterministic “Juice” budget probes and rolling behavioral probability fingerprints.
• Three preset tiers (low/medium/high) scale from 19 quick probes up to 158 requests across four distinct context shapes.
• Explicitly warns against false positives: official upstream risk controls, IP reputation, or temporary routing can trigger mismatches that look like vendor fraud.
• Licensed under PolyForm Noncommercial 1.0.0—source is visible, but commercial use requires separate written permission.

Caveats

• The README notes that even a clean bill of health cannot rule out transparent proxies, probe-aware routing, or differential treatment of test traffic versus production traffic.
• Custom probes and modified parameters drop into an unofficial “custom tier” that only shows reference match percentages, not hard conclusions.
• Native Codex request spoofing requires TUN-level networking or careful HTTP/mixed proxy setup; SOCKS-only or PAC-only environments are explicitly unsupported.

Verdict

Worth a spin if you are a researcher or consumer auditing a GPT-5.6 reseller and want evidence that stays on your own machine. Skip it if you are looking for an officially blessed, commercially licensed compliance tool—this is explicitly a non-commercial, DIY forensics kit.