How UberLLM works: where the capacity comes from
UberLLM is an API for people who buy AI: one key for text, image, video and music models. UberLLM does not run the models. Each request goes to a third-party provider that serves the model, and you pay what that provider charged. No UberLLM fee on requests. Adding credit by card costs 5.5% + $0.50.
Why can a price be 90% below list?
Each seller sets its own price as a discount off list price, and cheaper offers that work get more of the requests, so sellers compete on price. The partner market UberLLM connects to says some sellers are selling capacity that would otherwise expire or go unused, and others have discounted access to a provider. UberLLM cannot see a seller’s own costs or contracts. The price shown is what a request is actually billed, and the rules do not change with the price: Sellers must own or be authorized to resell the access they list under their provider’s terms; stolen, leaked, shared, free-trial and evaluation keys are prohibited; sellers are paid only for requests they actually served, after refunds and chargebacks; and a request that fails is not billed.
Where do Claude, GPT and Gemini requests come from?
These models are made by Anthropic, OpenAI, Google and xAI. On UberLLM, a request for one of them goes to a seller on the partner market, who passes it to the model through the seller’s own account with a provider. In the week to 2026-10-30, those requests ran through seller accounts at Amazon Bedrock, Anthropic, Google, OpenAI, OpenRouter, Snowflake, Venice AI, Bankr, Capminal. The partner market requires each seller to be authorized to resell that access under the provider’s terms, including any limit on reselling or on access by others, and suspends or removes listings when a provider reports a violation. It checks that each offer answers correctly and marks known provider addresses as verified; it does not prove which model produced an answer.
Who serves your request
A seller lists an offer: one model, served through the seller’s own account with a provider (an API key at that provider’s address), at a price the seller picks. Sellers either list directly on UberLLM or sell on a partner inference market that UberLLM connects to.
When you send a request, it goes to an offer for that model that is working right now, and cheaper offers get a larger share of requests. If that provider fails before it starts answering, the request is retried with the next one.
Providers that also sell these models directly, with the number of listed models each one sells:
- OpenRouter · 141
- Venice AI · 135
- Bankr · 62
- Morpheus · 57
- Amazon Bedrock · 52
- Fireworks AI · 27
- Mordiem · 26
- Together AI · 23
- Anthropic · 13
- Uncensored AI · 13
- Jatevo · 11
- Z.ai · 10
- OpenAI · 9
- Google · 6
- Groq · 4
- DeepSeek · 3
- Engy.ai · 3
- Mistral AI · 3
Each model page lists them for that model. A seller’s account can be at one of these providers or at another API host.
Why the price is often below list
Each seller sets its price as a discount off the model’s list price (the price you would pay the model maker directly). Routing sends more requests to cheaper offers that are working, so sellers compete on price.
The partner market says some sellers sell capacity that would otherwise expire or go unused, and others have discounted access to a provider. Some prices are 90% or more below list, and those are the prices requests are actually billed. Why can a price be 90% below list?
Prices move with the market. The price shown is the lowest offer right now; a request can cost more when the cheapest sellers are busy. For a hard limit, send provider.max_price (chat) or max_cost (image, video, music): the request only goes to offers at or under it and is never billed above it. Details: Pricing.
What sellers must agree to
- They must have the right to resell the capacity under their own provider’s terms (Terms, “Sellers”). If a provider’s terms do not allow reselling or access by others, its keys may not be listed. Stolen, leaked, shared, free-trial and evaluation keys are prohibited.
- They are paid only for requests their offer actually served, after any refunds or chargebacks on them.
- Their provider key is stored encrypted, used only to serve requests sent to their offer, and never shown again.
- Each model gets one test request before the offer is listed. An offer listed directly on UberLLM then gets buyers only once it passes our model check: test requests that look like normal traffic, compared with the real model on answers and token counts, repeated at random every few hours. An offer that fails is paused and its seller’s payouts are held.
- Health is measured on real requests. An offer whose requests fail is taken out of routing for a while, and longer each time it fails again.
Seller details: Sell capacity and the sellers guide.
What you pay
- No UberLLM fee on requests. Adding credit by card costs 5.5% + $0.50. The credit you get is the exact amount you pick.
- Each request is billed what its provider charged, from prepaid credit. Accounts start at $0.
- Requests that fail before the provider starts answering are not billed.
- Every request and its cost is listed in Activity in your dashboard.
- Credit does not expire while your account is open.
- Unused credit can be refunded on request within 30 days of purchase, minus credit already spent. Email [email protected] from the address on your account.
Your data
The provider that serves a request receives what you sent and handles it under its own terms. UberLLM saves each request and its answer for 90 days, then deletes them. We use them to fix failed requests, answer support questions and stop abuse, never to train models. The model, token counts, cost and time of each request stay on your bill. See the Privacy Policy.
FAQ
Why can a price be 90% below list?
Each seller sets its own price as a discount off list price, and cheaper offers that work get more of the requests, so sellers compete on price. The partner market UberLLM connects to says some sellers are selling capacity that would otherwise expire or go unused, and others have discounted access to a provider. UberLLM cannot see a seller’s own costs or contracts. The price shown is what a request is actually billed, and the rules do not change with the price: Sellers must own or be authorized to resell the access they list under their provider’s terms; stolen, leaked, shared, free-trial and evaluation keys are prohibited; sellers are paid only for requests they actually served, after refunds and chargebacks; and a request that fails is not billed.
Where do Claude, GPT and Gemini requests come from?
These models are made by Anthropic, OpenAI, Google and xAI. On UberLLM, a request for one of them goes to a seller on the partner market, who passes it to the model through the seller’s own account with a provider. In the week to 2026-10-30, those requests ran through seller accounts at Amazon Bedrock, Anthropic, Google, OpenAI, OpenRouter, Snowflake, Venice AI, Bankr, Capminal. The partner market requires each seller to be authorized to resell that access under the provider’s terms, including any limit on reselling or on access by others, and suspends or removes listings when a provider reports a violation. It checks that each offer answers correctly and marks known provider addresses as verified; it does not prove which model produced an answer.
Does UberLLM run the models itself?
No. Every request is sent to a third-party provider that serves the model, directly or through a market of providers. Each model page names the providers that also sell that model directly.
Where does the capacity come from?
From sellers. A seller lists a model it can serve through its own account with a provider (an API key at that provider’s address) and sets a price. Sellers either list directly on UberLLM or sell on a partner inference market that UberLLM connects to.
Are sellers allowed to resell it?
They must be. Under our Terms, anyone who lists capacity must have the right to resell it under their own provider’s terms. The partner market’s Acceptable Use Policy says the same: sellers must own or be properly authorized to resell access to the keys they list, and stolen, leaked, shared, free-trial or evaluation keys are prohibited.
Why is it often cheaper than going direct?
Each seller sets its own price as a discount off the model’s list price, and cheaper sellers that are working get a larger share of requests. A seller’s price has to cover its own cost, which depends on its setup. Prices move with the market, and a request can cost more when the cheapest sellers are busy.
Do I pay for failed requests?
No. A request that fails before the provider starts answering is retried with another seller, and a request that still fails is not billed.
Is there a fee?
No UberLLM fee on requests. Adding credit by card costs 5.5% + $0.50.
Does credit expire? Can I get a refund?
Credit does not expire while your account is open. Unused credit can be refunded on request within 30 days of purchase, minus credit already spent: email [email protected] from the address on your account.
Who sees my prompts?
The provider that serves the request receives what you sent and handles it under its own terms. UberLLM saves each request and its answer for 90 days, to fix failed requests, answer support questions and stop abuse, then deletes them. Only our own team can read them, and we never use them to train models.
Is UberLLM part of Uber?
No. UberLLM is not affiliated with Uber Technologies, Inc.
Questions: [email protected] · About UberLLM · UberLLM is not affiliated with Uber Technologies, Inc.