Two AI models did contractor-grade work on a real research assignment
Epoch AI asked six models to do its own research. Mapping a data center was the assignment where two could compete with a human contractor.

Epoch’s analysts know what a finished data-center entry looks like. It takes regulatory filings, news reports and satellite imagery to work out which buildings house the computers and what equipment they contain. On one Swedish campus, two Claude models delivered work the analyst judged comparable to a contractor’s first submission.1
The faster run was also the cheaper one
Opus 5.5 finished in under three hours at a lower cost than the typical 16-hour contractor assignment. Fable 5.2 took about six and a half hours, cost more than a contractor and supplied the stronger result. Both submissions still needed editing.1
Where this changes the job
A research manager could buy a first draft and spend analyst time checking the map. The crucial cost is how long an expert needs to find a plausible-looking mistake. A fast draft that forces a second analyst to redo the investigation offers a smaller saving.
The comparison comes from one organization’s assignments and quality bar. Across the full evaluation, none of the six models produced an output that met Epoch’s employee standard on any task.1