Effort

$1,100 of compute, and where it went.

Every model call in the study, priced at the providers' published API rates. Running the same research and benchmark yourself would cost about this much before counting your own time.

$1,100at API prices, all models

1.76 bntokens processed, 3.5 million of them written by the model

109agent-hours: 24 Claude Code sessions with their subagents, and 1,037 benchmark runs

1,040sources read, and 46 repositories reviewed from source

Spend

Claude priced per model at Anthropic's published rates, with cache reads and cache writes at their own rates; other models at their providers' rates.
Part of the work	Cost at API prices
Research, writing and building, in Claude Code sessions	$702
Benchmark runs of Claude Code	$332
Other models: graders, labellers and task writers	$53
Claude helper calls inside the benchmark	$13
Total	$1,100

Research

Item	Count
Sources read, cited or archived	1,040
Public repositories reviewed from source, each with a written verdict	46 (36 in full, 10 in part)
Research documents	34

Benchmark

Item	Count
Configurations specified as arms	107
Task bank versions frozen	17
Tasks in the final harm-check bank	69
Campaign runs	14
Benchmark runs of Claude Code	1,037
Tokens used by those runs	260,145,184
Agent-hours of those runs	27.1
Commits to the project and to the benchmark runner	118 + 64
Automated tests in the runner	356

Most tokens are cached context that the model re-reads on every turn; they are cheap per token but dominate the count. The Claude Code sessions did the reading, writing and analysis; the benchmark runs are Claude Code building the test tasks.