AI Desk · AI Fingerprinting
Compare
Pick two models. The bench shows whether their measured fingerprints agree, exactly where they part, and what their declared contracts share.
vs
Verdict
Tokenizer fingerprint
| Test string | A | B | Verdict |
|---|
Values are the marginal prompt-token cost of each string with the model's fixed overhead subtracted. An x marks a row excluded as corrupt for that model.
API surface
The declared serving contract, field by field| Field | A | B | Verdict |
|---|
Try a comparison
Each opens a worked example with its verdictOx Alpha and Z.ai: GLM 5.3the stealth model and its likely familyopenMicrosoft: Phi 4 and IBM: Granite 4.1 8Btwo labs, one shared open vocabularyopenStepFun: Step 3.7 Flash and DeepSeek: DeepSeek V3tokenizer reuse across labsopenZ.ai: GLM 5.3 and OpenAI: GPT-5 Nanotwo unrelated families divergingopen
Or choose two models above: the address updates so any comparison can be linked and cited.