A study found that Anthropic's AI model “Claude Opus 5.5” completed a data center research task in under 3 hours that outsourced workers typically take about 16 hours to finish.
According to blockchain media outlet CryptoSlate on Nov. 18, AI research group Epoch AI assigned the model a task to investigate “Eco Data Center 1” and register it in its database. It provided research guidelines and tools and had the model organize facility information based on media reports, filings submitted to regulators and satellite images.
Opus 5.5 received 87.5 percent in the evaluation. The cost of the work was also lower than using outsourced workers. “Claude Fable 5.2,” which performed the same task, scored higher at 97.5 percent, but took about 6 hours 30 minutes and cost more than using outsourced workers.
OpenAI's “GPT-6.2 Astra” scored 65 percent on the same task. According to Epoch AI, the research materials included evidence that there were multiple buildings for AI computing, but the model judged the evidence insufficient and included only one building in its report. The report also omitted the computing scale and the number of chips that could have been estimated from the materials.
Epoch AI's data center analyst assessed that drafts produced by Opus 5.5 and Fable 5.2 were at a level similar to what outsourced researchers submit. Both models met internal standards on key research items, including finding reliable sources and judging facility information in line with the evidence they secured. Results from outsourced workers also typically undergo revisions before publication.
Drafts produced by the two models also needed revisions before publication. Opus 5.5 had to correct the classification of some equipment and satellite image labels, while Fable 5.2 had to fix inaccurate satellite image labels and shorten overly long explanations.
Epoch AI evaluated a total of 6 models based on actual internal work, including data center research as well as graphic design, data analysis and writing research reports. A perfect score means a level at which an output can be published without edits. No output received a perfect score in this evaluation.
The evaluation is meaningful in that it compared AI models' research work capabilities by looking not only at performance but also output quality and cost. It also showed more clearly that scores and cost-effectiveness can differ even within the same model family, providing a clearer basis for decisions on use.