A daily ceiling, checked first
The budget is checked BEFORE a turn runs, not after it has finished. Checking afterwards is how a ceiling becomes a report.
Cost
$0.037 for onecandidate, endto end
Screen the market, pull the indicators, reason about the setup, return a typed verdict. 127,000 tokens and fourteen tool calls, measured on a real run rather than estimated from a price list.
Where the money goes
Tool calls, not the answer
The visible output of a turn is a paragraph. The cost is the fourteen round trips underneath it — candles, indicators, the instrument’s specifications, your position, the venue’s state — each one adding to the context the model then reasons over. A chat that never fetches anything is cheap and worth what it costs.
Context grows through a turn
Every tool result stays in the conversation for the rest of the turn, so the tenth call is more expensive than the first. That is why the loop has a step budget at all — not to constrain thinking, but because a runaway loop is a runaway bill.
A model ladder, not one model
Heavy analysis goes to a stronger model; routine lookups go to a cheaper one. Most of what an assistant does is routine, and paying frontier prices for "what is the current price" is how per-question costs get to a dollar.
A step limit tuned by an actual failure
The loop was capped at eight steps when a turn meant one symbol and one indicator. Two live turns died on the eighth step — money spent, the answer cut off mid-sentence — so the cap is sixteen. The real limiter is the token and dollar budget; the step count only shapes the loop.
What keeps it bounded
Cheap waiting
A "not yet" verdict names a price condition and a cheap poller waits for it. One universe fetch answers every pending condition at once, instead of one expensive turn per check.
Idempotent alerts
A condition that stays true does not re-trigger analysis. Most runaway costs in this class of product are a loop re-analysing the same unchanged fact.
Per-tenant caps
Spend is attributed per account, with ceilings and a fair-share limit, so one busy account cannot consume the platform’s capacity.
Questions about cost
Why publish this?
Because the alternative — "unlimited AI" — is a tier that ends abruptly and without warning, and because the number explains the pricing. What this platform is billed for is data and model calls, so that is what the plans ration.
Does a simple question cost the same?
No. A lookup is a fraction of a cent; creating a screen from a sentence has been measured at under a penny. The 3.7 cents is a full grounded analysis with a verdict.
Can I bring my own model key?
On the paid tiers, as an option. It is deliberately not the free tier’s entry requirement — a free tier that opens with "first create an account at a model provider" is a funnel with no entrance.
Spend four cents finding out
The free tier includes a budget of turns against live data.