configurator
Price your agent stack. Check our math.
Four choices, one receipt. Everything below is computed from the published rate card with the assumptions printed next to the answer. It is the same formula that generates the worked examples on the pricing page.
sample rates · fictional company
Estimated total at 100k runs/mo $384.00 · $0.00384/run
Your stack, priced
Customer support · Economy · Standard guardrails · Sampled traces
Metered per run
- Model tokens (10,800 in + 2,100 out)
- $0.00288
- Guardrail checks (2 per run)
- $0.00016
- Observability
- $0.00000
+ orchestration fee, volume-tiered below
Monthly, at three volumes
Swipe the table sideways to see every column.
| Runs / mo | Models | Guard | Obs | Orchestration | Total |
|---|---|---|---|---|---|
| 10,000 | $28.80 | $1.60 | $0.00 | $8.00 | $38.40 |
| 100,000 | $288.00 | $16.00 | $0.00 | $80.00 | $384.00 |
| 1,000,000 | $2,880.00 | $160.00 | $0.00 | $530.00 | $3,570.00 |
Effective cost per run at 100k: $0.00384
Assumptions: check our math
- Customer support: 6 model calls per run, averaging 1,800 input + 350 output tokens per call.
- Economy tokens at $0.15/M input, $0.60/M output. Illustrative list prices, passed through at cost.
- Guardrail checks run on our hosted screening model at $0.08 per 1,000 checks.
- Orchestration fee: $0.0008/run for the first 100k, $0.0005 for the next 900k, $0.0003 beyond.
- Worst case shown: provider prompt caching typically serves 60% to 90% of input tokens on multi-step runs at about 90% off, so your model line will usually come in under this.
Sample rates for a fictional company. Your selections save locally in this browser; this is a demo and nothing is sent anywhere.
reading the estimate
Three things the totals are telling you
Tokens dominate, until they don't
On frontier models the model line is about 95% of the bill and every prompt byte matters. On economy fleets at volume, guardrails and observability can cost more than inference. That flip is normal, and the controls are the point.
The worst case is printed
We assume zero cache hits. Real multi-step agents re-send accumulated context, and providers serve 60% to 90% of it from prompt cache at about 90% off. Your model line usually lands under the estimate, never over it from our side.
If the number scares you, good
A frontier research fleet at scale is genuinely expensive. Before you commit, read when not to use an agent. A pipeline might do the job for 6× less.