# Model authenticity check: which model is really behind a relay / third-party API

Source: https://modelbank.ai/verify · ModelBank · free · runs in your browser, API keys are never stored

## What it does
- Identifies which model an API endpoint actually routes to — even with a custom model name such as `tc-code-latest`.
- Detects relays that swap or downgrade the model, or inflate token counts.

## How
1. **Manual (no key)**: copy 3 prompts to the model and paste the replies back.
2. **API check**: enter base URL, key and model; about 1–3 minutes and 7–25 requests, billed to your own account.

## Evidence
- Output fingerprint: the model writes a few hundred integers on first instinct; each family has a stable "number handwriting".
- Tokenizer: identified from token usage (62 tokenizer families).
- Billing, hidden system prompts, returned model name, protocol details.

## Accuracy (held-out, 792 groups)
- Model family identified: 99.2%
- Exact model identified: 93.6%
- Cross-family impostors caught: 100.0%
- Genuine models wrongly flagged as inconsistent: 1.0% for one round, 0.1% after the API check's automatic second round

## Library
66 models in 15 families: claude, deepseek, gemini, glm, gpt, grok, hunyuan, kimi, mimo, minimax, muse, openpangu, qwen, seed, step.

## Privacy
Keys are used only in the browser; when a relay is needed requests are forwarded unchanged and never stored or logged.