What should AI work cost?
Community-reported task costs, benchmark-anchored pricing, and percentile spend budgets by role — so individuals stop guessing and executives can plan with real numbers.
Every task type and role can report — coding has the deepest benchmark anchors today, and role budgets are seeded from published references until community-reported cells reach n ≥ 5.
Find your starting point
TokensBench answers a different question depending on who is asking.
Am I doing it wrong?
See what people paid for tasks like yours, broken out by model, harness, and technique.
Ask what a task costs →What should we budget?
Per-seat monthly bands by role, sourced and exportable — build an AI budget you can defend.
Open the planner →How do we compare?
Cross-model, cross-harness cost data by role, checked against benchmark reference points.
Browse roles →Numbers worth knowing
A few headline figures from the references TokensBench is seeded on.
Ramp AI Index, June 2026.
DX benchmark, 2026.
Anthropic pricing page — tracked in our append-only price history.
How it works
Three layers, deterministic math — the LLM never computes a number.
Reference layer
Model pricing and benchmark cost-per-task, ingested from authoritative sources like SWE-bench, HAL, and Artificial Analysis — cited, versioned snapshots.
Community reports
Real task and period costs submitted by the community, moderated before publishing and weighted by verification.
Derived percentiles
Median, IQR, and p90 bands and budget plans, computed deterministically from the two layers above.