OpenRouter Pricing Explained: Tokens, Free Tiers, Fees
AI basics - - Owen Castellano
An LLM API bill has one unit: the token. Everything else, from router fees to free tiers, is a rule about how many tokens cost how much, or how many you may use.
Input, output, and why they differ
A token is a chunk of text, roughly a word or part of one. Providers count the tokens you send (input) and the tokens the model writes back (output), and price them separately per million tokens, written MTok. Output costs more. On Anthropic's pricing page, as of October 2026, Claude Sonnet 4.6 is $3 per MTok for input and $15 per MTok for output; cache hits, meaning input the provider has already stored, are $0.30 per MTok (Anthropic pricing). At that ratio of five to one, a chatty model can cost more than a long prompt.
The context window is the cap on input plus output for one request. It is a limit, not a price, but a long window invites long prompts, and those are billed in full on every call unless caching applies.
What a router adds
OpenRouter sits between an app and many providers behind one key. Its pricing page says it does not mark up inference, so token prices match the provider's list price. The fee is on buying credits: 5.5% on the Standard pay-as-you-go plan and 8% on Business, per the same page (as of October 2026, OpenRouter pricing). So the router cost is a percentage of the money loaded, not a number on each request.
A worked estimate
The sizes here are assumptions for the example, not measurements: one request of 2,000 input tokens and 500 output tokens, ten thousand requests a month, at the Claude Sonnet 4.6 list rates from Anthropic's pricing page quoted above, with OpenRouter's 5.5% credit fee on top.
| Step | Calculation | Result |
|---|---|---|
| Input cost per request | 2,000 / 1,000,000 x $3 | $0.006 |
| Output cost per request | 500 / 1,000,000 x $15 | $0.0075 |
| Per request | sum | $0.0135 |
| Per month | x 10,000 | $135 |
| Credits to load with a 5.5% fee | $135 x 1.055 | about $142.43 |
Rates and fee come from the Anthropic and OpenRouter pricing pages linked above.
The estimate is for list prices only. Change one assumption and the answer moves: shrink the prompt to 1,000 tokens and the input cost halves, while the output cost stays.
Free models and their limits
Free models on OpenRouter carry a :free variant name and a request limit instead of a price. The pricing page lists 20 requests per minute and 50 per day, with the daily cap rising to 1,000 once at least $10 of credits has been bought. Failed attempts count toward the daily cap, and a negative balance blocks free models until credits are added. For a prototype that is enough. For a feature with users, 50 requests is an afternoon.
What the price does not show
- Rate limits on paid models. They are a separate setting from price.
- Cache write costs. Cache hits are cheap, but writing the cache has its own rate, higher than plain input on the Anthropic page.
- Output length. The model decides how much to write, so a cap on output tokens is the only hard limit on cost per request.
- Queues and latency tiers on cheap or free routes.
Where it shows up in tool descriptions
Product pages rarely say "tokens". They say "unlimited messages", "credits", "resolutions" or "fair use", and each of those hides a conversion into tokens that the vendor does not publish. A free plan described this way is a request limit, not a price. The one description that avoids the question entirely is a model that runs on your own machine: the catalog's Faraday.dev (now Backyard AI) card describes a free desktop app that runs open models locally with no data sent to the cloud. There is no token bill there, but the hardware and the choice of models are yours to carry.