Engineering evidence · Public summary
AI Engineering Usage
Source-Data Reconciliation
Anthony W. Licausi · Prepared November 10, 2026
What was measured
A reproducible internal reconciliation of two usage exports: Claude Code daily aggregates and a Grok/Cursor event-level export. The common observation window was October 10–November 7, 2026. The calculated mean was 16,606,398.55 non-cache-read tokens per day.
Calculation method
For each reporting date: uncached input/prompt tokens + output/completion tokens + cache-creation/write tokens, summed across both exports. The 29 overlapping daily totals were divided by 29. Cache-read tokens were excluded from the headline figure.
Interpretation and limits
- This is an internal, reproducible source-data reconciliation, not independent third-party assurance or verification of provider metering.
- Claude Code uses Eastern Time date labels; Grok/Cursor event timestamps were grouped by UTC date. Dates were matched by label, without full timezone normalization.
- Only two available exports were included, not every model provider or AI product used.
- Cache writes may include repeated context. This is token-processing volume, not unique tokens, software output, quality, or a developer-percentile ranking.
- No claim is made about a Google conference percentile benchmark.
Verification
Original source exports, daily calculations, source hashes and a reproduction script are retained in a separate private evidence package and may be made available for appropriate confidential review. Raw event-level records are not publicly distributed because they may contain account identifiers, email addresses and internal project paths.