Loading…
The demo costs nothing. Then it meets a month of real traffic. Pick a scenario, choose an architecture, and watch what it bills: the number that decides a design is monthly, not per call.
What is simulated here is the arithmetic: tokens × rate × requests. What is NOT simulated is quality: no tier carries an accuracy score, because inventing one would be a made-up benchmark on a public page. Where capability matters, the scenario says so in the open and you can disagree with it. Rates are credits, not money: real prices move monthly and differ by vendor, but the shape does not, output costs several times input (1/4 versus 40/160 per million tokens, small against frontier).
Every incoming support message is sorted into one of six buckets and given a one-line summary. High volume, short text, and most of it is the same handful of questions.
Model tier
Credits per million tokens, input / output.
Cache
Share of requests answered without calling a model. The rates are this scenario's assumption about how often people ask the same thing: an exact-match cache is useless where everyone words it differently.
Context
Input tokens per request. Sending the whole document is the version that works on the first try.
Routing
Send the 80% of requests this scenario considers simple to the Small tier, and leave the rest with the tier you chose.
Monthly bill
4,032
credits, against a budget of 200
Not there yet: the bill, a broken requirement, or both.
Cost is the one you can play with in a browser. The rest (retrieval that finds the right passage, an agent that stops, a deploy that survives Monday) is the course, and the seventeen projects you finish in it.