Tokener CA is a service for talking to AI models. You create a key, then pay the way that suits you: as you go, with funds you add to your balance, or with a monthly Agent plan that includes usage up to a 5-hour and a weekly limit. The Agent plan guide explains how the monthly plan works.
This page explains the handful of words the rest of the docs use. If you already know them, skip to Quickstart without code.
The five ideas
Model — the AI that reads your message and writes a reply. Tokener offers a small, curated set; each one has its own strengths, speed and price. See the current list and prices on the models page.
Token — how text is measured. A token is roughly a few characters; a short sentence is a few dozen tokens. You are billed by the number of tokens in your input plus the number in the model's reply, so shorter prompts and replies cost less.
Balance — money in your Tokener wallet, held in US dollars. You add funds up front and each request draws from the balance. Nothing is charged until you make a request. Instead of paying from your balance, you can choose a monthly Agent plan, which pays for your requests up to a 5-hour and a weekly limit.
API key — the secret that lets your code (or a tool) use your account. It starts
with tok-. You create it once, copy it, and keep it private; anyone with
the key can spend your balance.
Protocol — a way of formatting a request. Tokener speaks three, so you can use the tools and libraries you already have: OpenAI-compatible (Chat Completions), OpenAI Responses, and Anthropic-compatible (Messages). They reach the same models and use the same one key and wallet — you pick whichever your tool expects.
What a request costs
You pay for the tokens in your request and the tokens in the reply, each at the chosen model's per-token rate. Some models also have a cheaper rate for cached input (text they have seen very recently), which is applied automatically where it exists. A request that fails is not charged. Developers who want the exact rule can read How the numbers are counted and billed.
The exact rates are per model and can change, so they are never printed here — the models page always shows the current numbers.
How we measure uptime
The models page, each model's own page and the featured models on the home page show Uptime with a percentage and the period it covers — usually "last 3 days", followed by small blocks, one per day, oldest first, each coloured by that day's result (today's block is marked "today so far"). Point at a block to see its figure. The figure is defined as:
Share of our automatic test requests in the last 3 UTC days (today so far and the two days before) that got a complete answer. We send them about every 30 minutes from our server to the suppliers your requests use by default; supplier timeouts and rate limits count as failures. A short outage between two checks can be missed.
Some models may show a shorter period — the last 24, 12 or 6 hours, with one block for the whole period — or the last 7 days. The card always names the period, and the definition is the same with that period ("in the last 6 hours", "in the last 7 UTC days (today so far and the six days before)").
It updates by itself. When there are too few test requests so far, the card says Not enough data yet instead of a number, and if our test requests stop for a while the figure is hidden rather than showing an old one.
Where your spending shows up
- Usage lists every request with its model, protocol, token counts and cost.
- Transactions has two views: Payments lists what you paid by card (top-ups and Agent plan charges, each plan charge with its invoice), and Balance activity shows your balance changes, with usage summed per day.
Next
Ready to try it? Follow Quickstart without code — from sign-up to your first reply, no programming required.