Pricing
Pay for tokens, nothing else
There are no plans to compare. You top up in your own currency and spend the balance on requests, at the rate of the model you chose.
Why one dollar is possible
Most card processors charge a percentage plus a flat fee, typically about 30 US cents. The flat part does not shrink with the payment: it is negligible on fifty dollars and roughly a third of one dollar. That is why a five or ten dollar minimum is normal.
Payment here is a percentage with no flat component, so a small top-up costs proportionally what a large one does. The $1.00 minimum is a consequence of that, not a promotion, and it is the processor's own floor rather than a number we chose.
What a request actually costs
Four ordinary jobs at today's rates, priced per thousand runs. One run of any of them costs a fraction of a cent, which is true but hard to plan against.
A short support reply
llama-3.1-8b · 250 tokens in, 400 out
$0.032
per 1,000 runs
Extract fields from a form
mistral-nemo · 800 tokens in, 120 out
$0.028
per 1,000 runs
Summarise a five-page document
gemma-3-27b · 3,000 tokens in, 500 out
$0.48
per 1,000 runs
Reason over a long thread
llama-3.3-70b · 12,000 tokens in, 900 out
$2.23
per 1,000 runs
Token counts are for English. The same job in Swahili or Amharic uses more tokens and therefore costs more. We measure and publish that per model.
Rates by model
Dollars per million tokens, cheapest first. Prices are set in USD because that is what the compute costs.
OpenAI whisper-large-v3-turbo 
Mistral AI mistral-nemo 
Meta llama-3.1-8b Mistral AI mistral-small-24b-2501 
Google gemma-3-4b Google gemma-4-e4b OpenAI gpt-oss-20b 
Microsoft phi-4 
Qwen (Alibaba) qwen3.5-9b Google gemma-3-12b Google gemma-3-27b What you are not charged for
No subscription
Nothing recurring, and nothing to cancel. Top up when you need to.
No expiry
Your balance stays on the account until you spend it.
No card on file
Pay with the methods available in your country. Nothing is stored to charge you later.
No minimum
A dollar is a valid top-up, and so is every amount above it.
One thing that can cost more than it looks
Some models think before they answer, and the thinking is billed. The same one-line reply can cost a hundred times more on a reasoning model than on an ordinary one, so the price per token is not the price per request. Every model that does this is marked on the models page and in the playground, and comparing them on your own workload is the only way to know.
You are billed for the tokens a request actually produces, and every response carries its own cost. The checkout page shows the exact total before you confirm, and you receive the full amount you buy.