DEVELOPER DOCUMENTATION

Pricing & free models

Know the benchmark. Know the price you can actually call.

Four prices, one token basis

• Direct list: the model’s direct price reported by the market source.
• Cheapest OpenRouter endpoint: of every active Standard OpenRouter endpoint for the model, the one with the lowest input + output cost for a typical request of that model (see below). Flex, batch and priority rates are excluded. Provider discounts are already included. If endpoint prices are unavailable, the page shows an explicitly labeled catalog reference and claims no discount. These are base token rates; long-context and request-specific charges can differ.
• Market price: the cheapest observed qualified offer for the selected token basis.
• Our price (estimated): the selected gateway’s current eligible route. When labeled “budget estimate,” it includes a routing price buffer and is a ceiling for budgeting; the final billed cost can be lower. The workload savings calculation uses this estimate when available. Market observations can be cheaper and do not guarantee execution at that price.

Discount calculation

A model’s discount is quoted for one typical request of that model, against the cheapest OpenRouter endpoint for the same request:

input = 1,000 × r tokens
output = 1,000 tokens
cost = input × input rate + output × output rate
discount = round((1 − our cost / cheapest OpenRouter endpoint cost) × 100)

r is the model’s input tokens per output token, measured on GPU Router production traffic, 7 days ending 2026-10-09. Models without their own measurement use r = 80. Our price is the best price available right now, including the routing fee: the cheapest of every eligible route across our suppliers and the live offers of trusted seller groups on market routes (on a discounted route, only offers under its price ceiling). Model pages, the workload calculator and the price chart use the same price; the calculator opens on a thousand typical requests. Cache pricing is left out on both sides: our suppliers pass the upstream cache discount through at the same ratio as the model’s own cache pricing, so the comparison is the same without it. A discount is shown only when it is positive and a current OpenRouter endpoint reference exists, as a whole percent and never above 95%.

Price history compares each recorded market price with today’s reference, because OpenRouter history is not recorded. The workload calculator compares your own token mix with the cheapest endpoint for that mix. Charts label current reference prices explicitly; they are not fabricated historical observations.

The 99.5% free threshold

Both input and output rates must be no more than 0.5% of the corresponding nonzero direct prices. Qualification uses exact micro-dollar amounts before display rounding. Unapproved, stale, missing-reference, and empty-capacity offers cannot qualify.

“Free candidate” means the market price meets the threshold. A request is free only when the gateway exposes an active sponsored route with zero customer charge. A candidate never silently becomes a paid request.

Admin control

Administrators enable free-model visibility and access in Settings → Admin. It starts disabled. Turning it off hides free listings and promotions and blocks new sponsored calls; requests already admitted may finish. Provider eligibility and spending limits always apply.

Calling a free model

Choose a model ID ending in :free from the gateway’s live catalog. Both customer token prices are $0. An API key is still required; an active stake or credit balance is not. The sponsor pays the provider using a separate daily allowance.

Free calls allow up to 2,048 output tokens, 16,384 estimated input tokens, and 100 attempts per wallet per UTC day, shared across its keys. A depleted sponsor allowance returns 429. A vanished qualifying route returns 404. Neither error switches to paid inference.

The sponsor reserves the executable route ceiling, which can exceed the observed qualifying floor because the venue uses whole-percent discount paths. That difference is funded by the sponsor. A trusted-provider allow-list stays attached to dispatch.

Settlement

Paid calls reserve credits before dispatch. Final provider receipts take precedence over token-price estimates. Missing usage or interrupted streams may require conservative settlement. Check the request receipt for its cost basis.

Failed and interrupted requests

A provider can bill a request that fails after it starts. When GPU Router retries the request on another route, you pay only for the attempt that answers. When nothing retries it (a stream you stop, an error or empty reply returned to you, a stream that stalls, or a request interrupted before it settles), you pay what the provider billed for that request plus the routing fee.

The failed request’s hold is released at once. Its charge appears on the request in Activity when the provider confirms the bill, usually within minutes. Requests the provider refuses before doing any work, never bills, or does not confirm within 6 hours are not charged.