Is your AI provider serving the model it claims? Test it now

Know which model you're paying for.

authenticated.si tells you if an AI model is real or fake. We check providers against the official weights, in seconds, for cents.

55 of 67
cheaper copies caught across 14 models
~2 s
for a live check
<$0.001
of GPU time per check

Which provider serves the real model?

Both claim Qwen2.5-1.5B-Instruct. One secretly serves a cheaper 4-bit copy.

  1. 1Ask both
  2. 2Pick the real one
  3. 3SPOT checks

Provider 1Qwen2.5-1.5B-Instruct

Ask a question below and Provider 1 answers here.

Provider 2Qwen2.5-1.5B-Instruct

Ask a question below and Provider 2 answers here.

Connecting...

3 free rounds a day

How it works

How do you know it's really
the model you chose?

You don't, until you check. SPOT checks every provider against the real model before your agent uses it.

  1. 1Run

    We download the official weights and run them on our own GPUs.

  2. 2Check

    We ask the provider the same secret questions and compare its answers with ours.

  3. 3Route

    Your requests go only to providers that pass, each answer with a signed receipt.

Proof of Model

Every verdict, signed and dated.

A snapshot shows which model a provider was serving, and when. These are the latest, straight from our GPU.

Time
When the check ran, to the minute.
Verdict
Genuine or Altered, with the score.
Signature
Signed by SPOT. Any later edit shows.

Latest snapshots

Live from our GPU

View all snapshots
Snapshot at (UTC)VerdictScoreTokensCost

Score: distance from the real model. Up to 1.00× is normal. Cost: GPU time for the check.

The problem

Cheap models, sold at a premium.

Search a marketplace for one open model and you find dozens of providers selling it. Some serve a cheaper copy. Today you re-prompt until the answer looks right, or pin a trusted provider and pay up to 13 times more.

Accuracy

55 of 67 fakes caught. No honest run flagged.

We catch compressed copies, swapped models and hidden system prompts. Across 14 open models and 67 substitutes, the real model passed every held-out run.

FP8 and 8-bit copies are often too close to the real model to catch.

  • A different, cheaper model17 of 17 caught
  • 4-bit compressed copy11 of 11 caught
  • Same model, setting changed (thinking, reasoning effort)9 of 9 caught
  • Hidden system prompt6 of 6 caught
  • "Uncensored" fine-tune posing as the original4 of 4 caught
  • FP8 compressed copy5 of 9 caught
  • FP8 attention cache2 of 8 caught
  • 8-bit integer copy1 of 3 caught
  • All fakes55 of 67 caught
  • Held-out honest runs flagged0 of 7

Who it's for

Pick providers on evidence.

Agent teams

Pick the cheapest provider without getting a cheaper model.

  • Requests go only to providers that pass
  • A signed receipt with every answer
  • Failing providers skipped automatically

Model labs

See how every provider serves your model.

  • Every provider checked the same way
  • Reference built from your release
  • A snapshot for each provider

Inference providers

Prove you serve the real weights.

  • Public snapshots of your endpoints
  • Your setup in the honest set
  • Disputes settled with evidence

Trust

Nobody can change a verdict. Not even us.

Every verdict says Genuine or Altered, and exactly how we got there.

Signed snapshots
Every verdict is signed and timestamped.
Fixed pass marks
Set from honest runs before any provider is checked.
No provider cooperation
Works with almost any provider serving any model, through its normal API.
Open method
Anyone can verify how we reached a verdict.

FAQ

Questions

Ask us anything at hello@authenticated.si.

What does authenticated.si do?
We tell you if an AI model is real or fake. SPOT runs the official model on our own GPUs, sends the same secret tests to a provider's API, and compares the pattern of answers.
Why would a provider serve a fake?
Open model weights are public, so anyone can sell access, and only the provider knows what actually runs. A cheaper copy under the same name saves them money. A March 2026 audit found 45.8% of endpoints weren't serving the model they claimed.
Can't I tell from the answers myself?
Not from one answer, and an agent picking the cheapest provider can't tell at all. Today most people re-ask the same question until it looks right, or pin a brand-name provider and pay 6 to 13 times more.
What do you catch?
Swapped models, compressed copies, and tampered settings like a hidden system prompt. In our tests every swapped model and changed setting was caught, and 4-bit copies 95% of the time.
What can't you catch yet?
8-bit copies and compressed memory sit very close to the real model, so we often miss them. When we can't tell, we say so instead of guessing.
Could you accuse an honest provider?
This is the hardest part, and we built around it. Pass marks come from running the real weights on many honest setups before we test anyone, and borderline results are re-checked with fresh questions. We never call a provider fake on a single check.
Does the provider have to cooperate?
No. SPOT calls the provider's public API like any other user, so it works with almost any provider of any open model. Hardware-based proofs need special chips at the provider, and only a few offer them.
If your method is open, can't providers game it?
Knowing the method doesn't help, because nobody can change the official weights we compare against. To pass, a provider has to answer like the real model across many questions, which means serving it. Only the questions stay secret, and checks come from ordinary accounts.
Why should I trust you?
You don't have to. Every weight file we run is matched to its official fingerprint, pass marks are fixed before any provider is checked, and the questions in each check are picked by a formula. The code is open source, so anyone can check the checker.
What if a provider disagrees with a verdict?
We reveal the exact question in dispute, with proof it was sealed before the check ran. The rest stay secret.
How much does it cost?
$0.25 a check, or $5 a month per model for hourly monitoring. A check takes seconds.
Have you checked real providers?
Not yet. Every result so far comes from 14 open models on our own GPUs, with honest and fake setups we built ourselves. Provider names in our demos are made up.
Do you see my prompts?
No. A check only sends our own test questions, so your prompts stay between you and your provider.
How do I get access?
We're opening a beta to a small group first. Leave your email at the bottom of this page and we'll write before launch.

Get the model you pay for.

We're bringing trust to open AI, one check at a time. Leave your email and we'll write before launch.