Model Purity Check | Detect AI API Relay Model Swapping
AIBench.cc checks model purity with response model fields, self-identity, tokenizer counts and protocol probes to detect model swapping, downgrade or quantization risk.
Quick Answer
A model purity check determines whether the model id you requested is likely backed by the intended model. AIBench compares response model fields, self-reported identity, tokenizer count deviations and protocol behavior, then marks inconsistent evidence as suspicious or downgraded.
How It Works
- 1Read model or modelVersion fields from the response.
- 2Ask the model to self-identify and match model-family keywords.
- 3Compare tokenizer counts against known baselines with a short prompt.
- 4Combine error codes, cache fields and usage schema into a risk grade.
Use Cases
- Suspect a relay replaced a large model with a smaller one.
- Validate whether a low-cost channel silently downgrades models.
- Filter opaque API providers before procurement.
- Investigate sudden output quality or style changes.
Core Metrics
model field
Model name declared by the response
Self-identity
How the model answers identity prompts
Tokenizer deviation
Whether token counts match expected ranges
Protocol behavior
Error codes, cache fields and usage schema
Model purity is not a single signal
Relays may use aliases, fallback, quantization or account pools. AIBench shows multiple evidence points because one signal alone can be misleading.
Why results are graded
Self-identity can be affected by system prompts, and tokenizer baselines vary. A graded result is more honest and easier to review than a binary claim.
Related Search Terms
FAQ
Can model purity checks prove model swapping with 100% certainty?
No. They provide strong risk signals and auditable evidence, but a remote probe cannot replace internal supplier audits.
Why can the same model name produce different purity results?
The channel may use aliases, fallback, account pools or rollout rules. Retest with fixed time, region and prompts.
Related Guides and Comparisons
Start a Real API Check
Paste a temporary test key and run the same probes for latency, model purity, cache, rate limits and cost.