Overestimated demand.
A startup budgets for a launch that uses fewer tokens than expected. There is more paid-for usage than work to spend it on.
ABOUT GPU ROUTER / THE OPEN MARKET FOR INTELLIGENCE
Intelligence has
a market price.
Great models. Unused capacity. A better deal for the people putting it to work.
GPU Router connects your requests to suppliers offering model access at competitive prices, including discounted surplus capacity. We bring those offers together through one API.
THE OPPORTUNITY
Capacity, put back to work.
Capacity already paid forRoom to use
Used capacitySurplus capacity
Supplier capacity
GPU Router
Your next request
01 / WHERE THE SAVINGS COME FROM
Sometimes, the budget
is bigger than the workload.
Teams plan ahead. They buy usage, reserve capacity, and make commitments. Actual demand can turn out differently. Our suppliers and the marketplaces they work with can make that unused capacity available for less.
Plans that change.
A team switches models, pauses a project, or finishes a big workload. Capacity booked for the original plan becomes surplus.
Time left on the clock.
Eligible prepaid usage or capacity may be nearing the end of its term. Selling access at a discount can recover value before it goes unused.
These are examples of how surplus arises. Available supply depends on each supplier’s arrangements and changes over time.
WHAT WE BELIEVE
A token is worth what
people will trade it for.
We believe there is a fair market price for intelligence, set by willing buyers and sellers. That price can be below a lab’s published rate. Making those offers easier to find gives builders more choice—and useful capacity another customer.
02 / OUR PART IN THE PROCESS
We’re the router.
Suppliers provide access to the models. We handle the connection, route selection, and usage billing. You choose the model and the privacy requirement for your request.
Choose your model.
Find the model you want and select Standard or ZDR.
We compare the routes.
Current supplier quotes, availability, and your request’s requirements determine which routes are eligible.
Put the capacity to work.
Your request goes to the lowest-cost eligible supplier to maximize savings. Usage is metered, and the model’s response comes back through the same API.
Prices you can compare. Catalog quotes include the standard routing fee. Discounts compare those quotes with the cheapest OpenRouter endpoint for a typical request of each model. Prices and supply change; your token mix determines the saving. See the pricing methodology ↗
PROVIDER CHECKS
Know how your route is checked.
We use explicit model mappings, restrict eligible marketplace suppliers, and test selected routes with context-dependent conversations, tools, and images. Those checks help us identify compatibility problems and exclude unreliable offers.
Coverage varies by model and supplier. Our checks do not independently prove model weights or certify a supplier’s business or resale rights.
Read our provider verification process ↗03 / YOUR DATA, YOUR ROUTE
Set the boundary.
Then find the price.
Zero data retention, or ZDR, restricts a request to serving routes whose provider policies do not retain prompt and response content. Eligibility belongs to the route serving the model.
FOLLOW A REQUEST
One model. Different eligible routes.
Your requestSame model.
Your privacy choice.
Your privacy choice.
GPU RouterZDR routes only
zdr: true- Supplier ALowest quote · Standard retentionExcluded
- Supplier BZDR-eligible quoteSelected
- Supplier CRetention unconfirmedExcluded
Privacy requirement applied. Then prices compared.Only the qualifying route receives the request. If none qualify, the request fails.
The policy stays with the request.
We apply the ZDR requirement before routing and keep it on retries. If no eligible route is available, the request fails instead of dropping the requirement.
Usage records still exist.
Model names, token counts, cost, and timing can be recorded for billing and operations. Our ordinary usage ledger does not store prompt or response bodies, and we do not train models on your prompts or outputs.
Check the full path.
The serving provider processes your content. Its policy may allow temporary memory caching. Your app, external tools, and stored or deferred workflows have their own data handling. See where your prompt goes.
A smaller eligible supply pool can mean a different price. Discounted ZDR routes are available when qualifying suppliers offer them. Compare the current quotes for your model.
Explore ZDR modelsMORE CAPACITY. MORE POSSIBILITY.