The site about AI costs shows you its own bill
We ask the community to report what their AI work costs, so we report ours too. This page tracks TokensBench build spend and, eventually, runtime spend, logged as it happens.
Build spend by phase
One row per sub-agent run, as reported by the Claude Code harness.
| Phase | Agent | Model | Tokens |
|---|---|---|---|
| design research | benchmark-sources researcher | claude-sonnet-5 | 69,977 |
| design research | competitive-landscape researcher | claude-sonnet-5 | 87,964 |
| design research | receipt-formats researcher | claude-sonnet-5 | 83,446 |
| build wave 1 | shell/home/methodology builder | claude-sonnet-5 | 162,896 |
| build wave 1 | core test-suite builder | claude-sonnet-5 | 146,535 |
| build wave 1 | submit/api/admin builder | claude-sonnet-5 | 240,583 |
| build wave 1 | planner builder | claude-sonnet-5 | 235,641 |
| build wave 1 | ask/cells/roles/models/benchmarks builder | claude-sonnet-5 | 254,632 |
| launch prep | launch-hygiene builder (health, sitemap, OG, privacy) | claude-sonnet-5 | 177,453 |
| launch prep | reference ingest pipeline builder (5 sources) | claude-sonnet-5 | 286,755 |
| launch prep | Artificial Analysis API adapter (same agent, resumed) | claude-sonnet-5 | 86,857 |
| launch prep | reporter withdrawal feature builder | claude-sonnet-5 | 155,023 |
| Total | 1,987,762 | ||
TokensBench's own build spend, logged as it happens — the site's first case study. Figures are agent-token totals reported by the Claude Code harness per sub-agent run (input+output). Orchestrator (Fable) session tokens are accounted separately at the end of the build. Updated 2026-07-31.
Runtime ledger
The runtime LLM ledger — tokens the live site spends classifying free-text /ask queries — is empty. v1 ships with the classifier dark (ASK_CLASSIFIER=off), so there is nothing to log yet. Once it flips on, per-category LlmUsage entries will appear here alongside the build log.