Pick a real public dataset, or upload your own trace export, and see a quarter of weekly AI bills bounce around, alongside the flat number we'd write to guarantee certainty. Everything runs in your browser, so your data never leaves this page.
A CSV with a header row and one row per unit, or one row per call. Costs must already be in US dollars, and tokens are not read.
cost_usd, cost, total_cost, total_cost_usd, usd or cost_per_unitunit_id, task_id, instance_id, trace_id, session_id or ticket_id. With one, rows are added up per id, so a per-call export prices correctly. Without one, every row counts as a unit.task_id,cost_usd,timestamp
tkt_1041,0.38,2026-08-02T14:11:09Z
tkt_1042,1.94,2026-08-02T14:12:31ZDownload a 60-unit sample CSVA quarter of weekly bills (same workload, two ways):weeks simulated from the dataset's own tasks with week-to-week mix shift
If your workload or appetite calls for something else, the same distribution prices these contract shapes:
One flat price makes easy units subsidize hard ones. If your task scope varies widely, we can quote Small/Medium/Large instead, each unit assigned a tier at intake, each tier priced as its own book.
| Tier | Share | Mean | Price |
|---|---|---|---|
| Small | 50% | $0.52 | $0.56 |
| Medium | 40% | $1.02 | $1.10 |
| Large | 10% | $1.76 | $1.89 |
Tiers cut at the p50/p90 of your own distribution. Live tiers come from task features at intake.
Not ready for a fixed price? Second Moment passes through costs up to an agreed threshold (plus a small fee) and absorbs anything above. This is where most contracts will start.
| Threshold | Fee / unit |
|---|---|
| $1.46 (p90) | $0.0372 |
| $1.69 (p95) | $0.0156 |
| $2.10 (p99) | $0.0024 |
Example thresholds set at the p90/p95/p99 of your own units. Fee = 1.25 × the expected amount we absorb above the threshold.
Pay only for resolved units, for agents that have a well-defined terminal state. It needs your resolution rate measured on a carried book first, and the price carries the failure and correlation load (roughly 1.4–1.8× the per-unit rate), so it fits workloads whose value per outcome clears it. Cap → fixed per unit → fixed per outcome. Each step naturally funds the data that prices the next.
mean $0.84 + risk load $0.02 (weekly aggregate q99 at 2,145 units) + 8.1% load = 4% base + 4.1% variance-linked (σ/μ 0.51) · Coding agent · 6,306 issues (SWE-rebench, public)
Your cost distribution n=6,306 · Z=0.97
| Unit cost | USD | × the average |
|---|---|---|
| the average unit | $0.84 | 1.0× |
| p95 | $1.69 | 2.0× |
| p99 | $2.10 | 2.5× |
| the worst unit in the sample | $2.84 | 3.4× |
| the fixed price | $0.94 | 1.1× |
Backtest: 13 weeks, blue we earn, red we absorb