Claude AI Completes 16-Hour Research Task in Less Than Three Hours
Claude AI finishes research task in record time
Epoch AI reported that an artificial intelligence model from Anthropic, called Claude, successfully finished a complex research job in less than three hours. This specific task usually takes human contractors about 16 hours to complete.
The test involved gathering data on a facility called EcoDataCenter 1. To do this, the AI had to look through news reports, government filings, and satellite pictures to find specific details about the building and its equipment.
Key performance findings
- Claude Opus 5.5 finished the job in under three hours and scored 87.5% on accuracy.
- Claude Fable 5.2 scored a higher 97.5% but took about six and a half hours and cost more than a human worker.
- A human contractor typically takes 16 hours to finish the same research.
- OpenAI’s GPT-6.2 Astra model scored 65% on the same task, missing several important details.
How Epoch AI tested the models
The research group gave the AI models the same tools and instructions used by their professional human researchers. They then graded the AI's work using the same standards they use for people. While the AI models were very fast, they still made some small mistakes. Humans had to step in to correct technical descriptions and image labels before the work could be used.
What has been confirmed
Epoch AI confirmed that Claude Opus 5.5 could perform research at a level similar to a professional contractor but at a lower cost. Both the Opus 5.5 and Fable 5.2 models met the organization's standards for core research, even though they were not perfect.
What remains unclear
It is not yet known exactly how much human time is required to review and fix the AI's mistakes before the research is ready for public use. The report also did not list the specific dollar amounts for the costs mentioned in the comparison.
How this changes human research roles
Because the AI is so fast at gathering information, Epoch AI is thinking about changing how it uses human workers. Instead of having contractors spend hours searching for data, the organization may have them focus on checking the work the AI produces. Humans would act as editors to ensure the AI's research is accurate and complete.