Do you get the model you asked for?
Yes, and we check. A seller on UberLLM can list only through a provider on our list, such as OpenRouter, Groq or Amazon Bedrock, with its own key there, so the provider runs the model, not the seller. Every offer is then tested and reviewed by our team before any buyer reaches it, and keeps being tested on live requests. A seller caught serving a different model is banned, and the buyers it served are refunded from its unpaid earnings. No one can yet prove which exact copy of a model wrote an answer, so this page also says what our checks cannot catch.
- 292
- Seller offers tested
- 0
- Passed and serving buyers
- 134
- Failed and stopped
- 3
- Sellers banned
- 79,555
- Test requests in the last 7 days
Updated November 5, 2026.
Where a seller’s offer runs
- A seller can list only through a provider on our list: real AI providers such as OpenRouter, Groq, Together AI, Fireworks AI, DeepSeek, Mistral AI, Z.ai, Alibaba DashScope, Meta and Amazon Bedrock. A seller’s own server can’t be listed.
- Your request runs on that provider’s own servers, with the seller’s own key there. The provider runs the model, not the seller.
- Each provider on the list is marked verified or not. Your requests reach only offers on verified providers, unless your key is set to also use unverified ones.
- A seller can ask us to add a provider. We add it only after testing it.
Providers sellers can list through:
- AkashML · Verified
- Alibaba DashScope — China · Verified
- Amazon Bedrock · Verified
- Bankr · Verified
- Capminal Gateway · Verified
- Concentrate.ai · Verified
- DeepSeek · Verified
- Engy.ai · Verified
- Fireworks AI · Verified
- Groq · Verified
- Meta Model API · Verified
- Mistral AI · Verified
- Mordiem · Verified
- OpenRouter · Verified
- Together AI · Verified
- TypeSafe · Verified
- Uncensored AI · Verified
- Venice AI · Verified
- Z.ai · Verified
- Z.ai Coding · Verified
- 0G Compute Router · Not verified
- Aiflow / Arcflow · Not verified
- Baseten · Not verified
- BlazeAPI · Not verified
- BytePlus ModelArk · Not verified
- CheaperInference · Not verified
- Dolphin · Not verified
- Dot · Not verified
- InferHub · Not verified
- Jatevo · Not verified
- Jatevo API Gateway · Not verified
- Kairence (NEURON) · Not verified
- Moonshot AI / Kimi · Not verified
- Morpheus · Not verified
- NaraRouter · Not verified
- NeuralWatt · Not verified
- Ollama Cloud · Not verified
- OneAPI · Not verified
- Orbio · Not verified
- Pareto Inference · Not verified
- SoftForge · Not verified
Before a seller gets any buyers
- The provider’s name for the model is checked against the model listed: it must be in that provider’s own model list and be the same model. A seller can’t list Claude and point it at a cheaper model.
- A test request when the offer is listed. An offer that doesn’t answer is never listed.
- An entrance test of about 55 requests that look like normal use. We compare how the offer counts tokens, who it says it is and how it answers questions that have one right answer with a trusted reference for the same model.
- A trial on copies of real requests: the seller answers a copy, the buyer gets the answer from an offer already checked, and we compare the two. Buyers are not affected.
- Probation for at least 7 days on a small share of requests (at most 1 in 10), with some of its answers compared again.
- A person on our team reviews every offer. No buyer reaches an offer before that review.
- A new key, provider or model counts as a new offer and goes through all of this again.
On live requests
- The model named in every reply is compared with the model listed.
- Test requests keep coming for as long as an offer is listed. They are worded fresh each time and travel the same path as your requests, so they look like normal use.
- Every reply a seller sends to a buyer is also scored for signs of a different model or padded counts. A high score sends the offer to our team.
- An offer whose provider says the key is no longer valid is taken off automatically.
- An offer flagged twice is stopped at once, and its seller’s payouts are held.
- You are never billed for padded counts: the prompt tokens a seller can bill are capped close to our own count of your request.
What the checks show, and what they can’t
What they show
- The offer runs on a provider we know, not on a seller’s own machine, with a key that works.
- The provider’s model is the model listed, and the reply names it.
- It counts tokens the way the real model does, so a model from another family shows up.
- Nothing sizable is added to your request, such as hidden instructions.
- It names its real maker, and it answers questions with one right answer as well as the real model does.
- Its answers read like the real model’s answers, side by side with a trusted reference.
- You are not billed for padded counts.
What they can’t show
- Which exact copy of a model wrote an answer. No software check can prove it; only hardware attestation can, and UberLLM does not have it.
- That the provider itself serves what it says. We rely on the provider, and on our own tests of what comes back.
- That an open model runs at full precision. Some providers serve smaller 8-bit or 4-bit copies, and those can pass. Sellers must state the precision they serve, and a false statement gets them banned.
- That every reply is served the same way as our tests. That is why tests look like normal use and why live replies are scored too.
- The quality of any single reply. The checks are samples, not a promise about one answer.
This page is about sellers who list on UberLLM. Requests that go to the partner market UberLLM connects to follow that market’s own rules: where Claude, GPT and Gemini requests come from.
When a seller cheats
- The offer is stopped, and the seller can’t turn it back on.
- The seller’s payouts are held. A seller caught serving a different model is banned: the account is suspended and every offer stops.
- Earnings not yet paid out are used to refund the buyers it served.
- The same provider key, payout address and email can’t be used to list again.
On the first day of checks, October 31, 2026, 10 offers were stopped. One sold as Claude said it was “GPT-5 by OpenAI”. Others added about 4,400 tokens of hidden instructions to every request.
Your controls
From the widest to the strictest:
| Control | How to turn it on | What it does |
|---|---|---|
| Also use sellers on providers we have not verified | A setting on the key, off unless you turn it on | Can be cheaper. These providers are not on our verified list. Their offers still go through every test and our team’s review. |
| Verified providers only | On for every account. Nothing to set. | Requests reach only seller offers on verified providers that passed the tests and our team’s review. |
| Choose who serves you | provider.only or provider.deny in the request | Only the routes or providers you name, or never them. For example, "only": ["seller"] sends a request only to UberLLM’s own checked sellers. |
To cap the price as well, add provider.max_price (chat) or max_cost (image, video, music). It limits what you pay, not who answers.
Every routing option: Models & routing.
Think a reply came from the wrong model?
Email [email protected] with the reply’s X-Request-Id (the same id as its row in Activity) and what made you suspect it. We test that offer again. A seller caught serving a different model is banned, and the buyers it served are refunded from its unpaid earnings.
Why a working key is not enough
Checking that a provider is real, the key works and the model name matches catches a dead key or a wrong name. It does not catch a reseller adding its own hidden instructions, or a cheaper model behind the right name. That is why UberLLM also tests what answers: how it counts tokens, who it says it is, how it answers questions with one right answer, and how its writing compares with the real model.
How the checks work (for developers)
- Token arithmetic: two requests that differ by a fixed text. The difference in billed prompt tokens depends only on the model’s tokenizer, so a model from another family shows up, and a hidden prompt cancels out of the difference and is caught on its own in the shorter request.
- Identity: a reworded “who made you” question. An answer that names another maker or product, such as a coding tool’s persona, is a flag.
- Capability: freshly generated questions with answers we compute ourselves (long multiplication, counting letters, date arithmetic, sorting, base conversion). The pass rate is compared with a trusted reference for the same model using a one-sided Fisher exact test.
- Output distribution: short completions of the same prompt, compared with the trusted reference using an MMD two-sample test (Gao, Liang and Guestrin, ICLR 2025).
- Tests are sent through UberLLM’s own public API on the same path as buyer requests, with fresh wording each time.
- A reference answer is used only if that reference call itself passed the token and hidden-prompt checks.
- Padding cap: the prompt tokens a seller can bill are capped near our own count of the request, with an allowance for images, tools and the model’s built-in prompt. Tokens above the cap are neither billed to you nor paid to the seller.
- Open models: sellers state the precision they serve (full, 8-bit or below 8-bit), and model pages show it.
FAQ
Does UberLLM check that I get the model I asked for?
Yes. Sellers can list only through providers on our list, and the provider’s model must be the model listed. Every offer is tested and reviewed by our team before any buyer reaches it, and keeps being tested on live requests. A seller caught serving a different model is banned, and the buyers it served are refunded from its unpaid earnings.
Can a seller run the model on its own server?
No. A seller can list only through a provider on our list, such as OpenRouter, Groq, Together AI or Amazon Bedrock, with its own key there. The request runs on that provider’s servers, so the provider runs the model. A seller can ask us to add a provider, and we add it only after testing it.
Can UberLLM prove which model wrote an answer?
Not with certainty: only hardware attestation can prove it, and UberLLM does not have it. The tests catch a different model, hidden instructions and padded token counts. A lower-precision copy of an open model, or a provider that itself serves something else in a way the tests don’t catch, can still get through, which is why live replies are scored too.
Can I use only verified providers?
That is the default. Requests reach only seller offers on verified providers, unless you turn on “Also use sellers on providers we have not verified” for a key.
Is a very low price a sign of a fake model?
Not by itself. Every offer goes through the same tests whatever its price. Sellers set their own discount, and a lower price does not change the rules or the checks.
Questions: [email protected] · How UberLLM works