You sell the outcome at a fixed price. The tokens are variable.
SixDecimal makes the variable part safe and legible — a hard ceiling and an exact cost, per client. Here is what that looks like across the three things agencies, studios and product teams actually resell.
You host the bot on a retainer. The traffic doesn't care what you quoted.
Support and sales bots you run for clients bill a flat monthly fee. Then one account gets a press hit — or one user decides to abuse the widget — and its token bill triples while your invoice stays exactly the same.
One account runs hot and you eat the overage.
A viral spike or an abusive user pushes a single client's bot far past what you priced. The provider meter keeps climbing; your retainer doesn't. That gap is pure loss — and you won't see it until the invoice lands.
A hard per-client budget refuses at the ceiling.
Set a daily or monthly budget per client. When that account reaches its ceiling the gate returns a fail-closed 402 — that bot degrades gracefully instead of quietly spending your margin, and every other client keeps running untouched.
Cost-per-client, read next to the retainer.
See each bot's real monthly cost beside what you charge for it. The accounts priced below cost stop hiding inside one lump provider bill — you reprice them at renewal instead of absorbing them another quarter.
Example figures. Cost is metered to the micro-dollar; the retainer is what you invoice. harbor-desk hit its cap and still costs more than it pays — a reprice at renewal, not a surprise in the invoice.
An agent in a loop spends at machine speed.
Multi-step agents and copilots call tools in a loop. Most runs are cheap. But a retry storm or a stuck plan can iterate thousands of times before anyone notices — on one client's key, overnight, against a price you already fixed.
One runaway burns a month of margin by morning.
A copilot stuck re-planning, or a tool that keeps 500-ing into a retry loop — the agent doesn't get tired. A single runaway can burn on the order of 1000× a normal run, and by the time an alert fires on the provider dashboard the hundreds of dollars are already gone.
reserve → commit stops the loop with a 402.
Every iteration reserves its estimated cost before the upstream call. The moment the next reserve would cross the client's cap, the gate refuses — the loop gets a 402 BUDGET_EXCEEDED and halts before the money leaves your account, not after.
See the true cost of “agentic”, per client and feature.
Attribution tags every step to a client and a feature — by header (x-sd-feature) or virtual key — so “agentic” stops being a line item you can't explain. You learn which agent features are cheap to run and which quietly need a higher price.
Example figures, projected. The cap held: the loop got a clean 402 and stopped one reserve short of the ceiling.
A schedule bug runs all night at ten times the volume.
Scheduled and batch automations run LLM calls over a set of inputs on a timer. When the input set is wrong or a cron fires twice, the job doesn't crash — it just processes 10× the rows and sends 10× the tokens, silently, until the run finishes.
A bad input set silently 10×'s the run.
Nothing errors. The job dutifully processes a duplicated or bloated input set, or a schedule bug re-fires it, and the cost of one run multiplies without a single failure to page you. You find out when the monthly meter is already spent.
A total or rolling budget bounds the blast radius.
Put a total or rolling_24h budget on the automation. When the run crosses it, further calls get a 402 and the rest of the batch is deferred — the damage is capped at the ceiling you chose, not the size of a bad input set.
Per-project cost, against the project price.
Attribute each automation to its project and read what the run actually cost against what you billed for the project. A workflow that erodes its own project fee shows up before you scope the next one.
Example figures. A duplicated input set 10×'d the run; the total budget stopped it at the ceiling instead of letting the whole batch through.
Attribute, cap, margin — per client, from day one.
Three use cases, one instrument. Under all of them SixDecimal runs the same three moves — the product's own reserve → commit → margin arc, applied to your whole book of clients.
Every token is tagged to a client, feature and team as it happens — by request header or virtual key, with no app rewrite.
A hard budget lives in the request path. The runaway call gets a fail-closed 402 before the spend leaves your account — not a warning you read tomorrow.
Attributed cost is joined to what you actually invoiced, per account, so a client that turned unprofitable is visible before the quarter closes.
Put a hard ceiling on every client's spend — before the invoice does it for you.
We're pre-revenue and onboarding a handful of agencies and studios as design partners. Attribute cost per client, cap runaway agents, and see real margin per account.