Claude Opus 5.5 Completes Epoch AI Data Center Task in Under Three Hours
Opus scored 87.5% on the assignment and cost less than a contractor. Claude Fable 5.2 scored 97.5% but took longer and cost more.

Anthropic’s Claude Opus 5.5 completed a guided data center database task in just under three hours, well below the roughly 16 hours contractors typically spend per data center.
The EcoDataCenter 1 assignment was one task in a broader evaluation of work drawn from Epoch AI’s operations. A Nov. 17 report described the evaluation; that date is the report’s publication date, not an established date for the model run. The comparison was not a controlled trial in which a model and a contractor completed the same assignment under matched conditions.
The models received a detailed guide and research tools. For the database entry, they reviewed news and regulatory documents, ordered satellite imagery and analyzed the images. The result applies to a guided assignment for one facility and does not establish performance across all data center research.
Opus scored 87.5% and cost less than a contractor. Claude Fable 5.2 earned 97.5%, took about 6.5 hours and cost more than a contractor. The evaluation did not provide dollar amounts for either model’s run. A perfect score would mean an entry could be published without changes.
Both models’ results were rated comparable to contractor work for the core research. Opus needed corrections to equipment classifications and more substantial revisions to its image annotations. Fable’s image annotations were imperfect, and its text was too long. The 16-hour figure is typical contractor labor per data center, rather than the result of a contractor completing this specific assignment in a controlled comparison.
The wider evaluation covered six models and multiple tasks. None met the employee standard on any task. That result puts the data center assignment in context: Opus produced work judged comparable to contractor output on this task, while the broader assessment did not find any model met Epoch AI’s employee bar across its tested work.
“We are now considering shifting contractors’ time toward reviewing AI outputs,” Epoch AI wrote in the report.