modelprint
One page that answers one question: who is really behind that API?
Run infrastructure probes against any OpenAI-compatible endpoint and compare
the fingerprints, side by side. Plumbing does not lie; personality does.
Try it: modelprint.ai
No install, no build, no server. Your API keys stay in your browser tab; every
call goes from your browser straight to the provider.
Why I built this
On 21 August 2026 a stealth model called Ox Alpha appeared on OpenRouter and
the whole timeline turned into a detective game. The best evidence was not in
anyone's hot take. It was scattered across replies: one person compared token
counts, one person read error codes, one person sent temperature 2.0 and let
the upstream error speak. Each trick lived and died in a single reply.
I collected them into one tool. And the same tricks matter far beyond stealth
models: two days earlier I showed that DeepSeek's API silently serves a
different model under the old deepseek-chat name. Providers move models
behind aliases and nothing tells you. This page is how you check.
What it found on day one
The full suspect lineup, 12 models, 9 probes each, one command:
mystery model (stealth/ox-alpha) against the crowd's guesses:
6/9 tokenizer 4/4 z-ai/glm-5.3
5/9 tokenizer 4/4 z-ai/glm-4.7-flash
2/9 tokenizer 2/4 openai/gpt-5.6-luna
2/9 tokenizer 1/4 qwen/qwen3.7-flash
2/9 tokenizer 0/4 kimi-k3, deepseek-v4-flash, minimax-m3
1/9 tokenizer 0/4 google/gemini-3.7-flash
0/9 tokenizer 0/4 x-ai/grok-4.6, claude-opus-5
Only the GLM family matches all four normalized tokenizer counts. Every other
lab's best is 2 of 4.
The probes
The approved registry currently contains 15 probes. These are the probes the
page loads and compares:
| probe |
what it reads |
why it identifies a lab |
| english pangram |
usage.prompt_tokens for a pinned text |
tokenizers are built per lab |
| chinese paragraph |
same, CJK text |
CJK segmentation differs most |
| code snippet |
same, source code |
indentation and symbol handling |
| emoji + rare unicode |
same, hard codepoints |
byte-fallback behaviour |
| template offset |
prompt-token overhead across two pinned short prompts |
the serving template's size |
| temperature: 2.0 |
the validation prose, verbatim |
written by the lab's engineers |
| max_tokens: 10^9 |
the refusal message |
it names the real output limit |
| error code family |
numeric vs string vs typed errors |
GLM's 1301 code gave Ox away |
| finish vocabulary |
finish_reason values |
vocabularies differ per lab |
Community batch: network forensics + capability + deep signals
| probe |
what it reads |
why it identifies a lab |
| router region |
the router's opt-in metadata snapshot (region, provider, strategy) |
the routing layer names who answered — a "Stealth" lane served by Z.AI ends one argument |
| generation record |
OpenRouter's /generation ledger for one call id |
provider name, data region and NATIVE token counts, straight from the router's books |
| header dna |
response header families (cf-ray, x-amzn-requestid, openai-processing-ms…) and the response-id prefix |
serving stacks expose different headers; Bedrock ≠ Vertex ≠ first-party |
| path split |
which upstream provider answers, and which payload it refuses — timing kept in the JSON export |
providers validate different parameters at different depths; the raw record also carries a GPU-free router→PROVIDER round trip, since both calls are refused and your own leg cancels out of the difference |
| context ceiling |
bisected accepted prompt-size bucket |
64k, 128k, 200k and 256k+ window classes help date the variant |
| logprob geometry |
normalized 3rd top-logprob gap δ on pinned continuations |
EVT predicts δ≈0.32 universally; deviations flag quantized or substituted weights |
Held for follow-up
Five proposals from the original community batch are not in the approved
registry yet: cutoff dating, wrapper leak, reasoning trace, stream cadence,
and the one-token battery. They remain follow-up ideas until they pass the
project's security and determinism gates.
Telemetry probes degrade honestly when a harness lacks the signal
(harness-lacks-http, headers-hidden-by-cors, logprobs-unsupported) —
an honest absence beats a fake fingerprint.
Tokenizer counts are normalized against a one-character baseline, so a
host's hidden template cancels out. Two hosts wrapping the same tokenizer
match; raw counts never would. Every tokenizer probe also runs twice and
reports unstable when a router spreads calls across different hosts,
instead of reporting noise as a fingerprint.
Why no censorship or personality probes: the community ran both on Ox
Alpha and reached opposite verdicts on the same day. Behaviour bends to a
system prompt. Tokenizers and error handlers do not.
Write your own probe
Every probe is one file in probes/ with the same contract:
// name: your probe name
// description: what it measures and why it identifies a lab
// author: your-handle
// version: 1.0.0
export const meta = { id: "...", name: "...", group: "tokenizer|errors|shape",
why: "...", long: false, author: "...", version: "1.0.0" };
export async function probe(ctx) {
const res = await ctx.chat({ messages: [...], max_tokens: 8 });
return { value: res.usage.prompt_tokens }; // the comparable fingerprint
}
Copy probes/_template.js, keep the rules in its
header (deterministic value, never throw, no timestamps), open a pull
request. Approved probes are one line in probes/index.js,
and your name goes in CONTRIBUTORS.md and on the page.
Thanks to @ItIsCuthNotCup,
@pjperez and
@toby-bridges, who wrote probes, fixed bugs
and corrected the documentation in the first days.
Honest limits
- Through a router, some answers are the router's, not the lab's. The tool
detects and labels these ("router-reported", "router-validated") and reads
the upstream error body whenever the router passes it through.
- A pinned host removes routing noise; the page fetches each model's real
host list and offers pinning for OpenRouter lanes.
- Matching fingerprints prove shared infrastructure, not identity. A lab can
serve two different models on the same stack. Read the tokenizer rows as
strong evidence about the family, not a signed confession.
Run the test suites yourself
node selftest.mjs # no key needed: every probe against a mock harness
node smoke.mjs # every probe against two real models, prints values
node suspects.mjs # the full 12-model lineup, prints the ranking
selftest.mjs runs everywhere; the other two need an OpenRouter key in the
environment they read.
Built by @unclecode, author of
Crawl4AI
If this is useful, follow @unclecode on X for the
next tool. MIT license: use it, change it, no need to ask.