1 min read
Add as a preferred source on Google

Epoch AI Finds Top Models Reached About 65% of Employee Work Quality

Claude Fable 5.1 and GPT-6 Astra shared the top position; models did better on structured work than open-ended assignments requiring judgment.

A sculptural scoring bar balances design, data and research work samples / TokenPost.ai
A sculptural scoring bar balances design, data and research work samples / TokenPost.ai

Epoch AI’s October evaluation found its top language models produced work averaging about two-thirds of the quality expected from employees, while open-ended tasks requiring judgment remained difficult.

Claude Fable 5.1 and GPT-6 Astra shared the top position in the evaluation.

Released Oct. 8, Epoch AI’s evaluation focused on work drawn from inside the organization.

The assignments included graphic design and research design.

Human graders scored each output against Epoch AI’s internal standards. A score of 100% represented the quality expected from an Epoch employee.

Claude Fable 5.1 and GPT-6 Astra both averaged about 65% across the assignments.

Models did better on structured work, including coding and data analysis. They had more difficulty with open-ended assignments requiring judgment or work tailored to a particular audience.

As of Nov. 17, the rankings remained unchanged from the October release.

Simon Yoon

Reporter

Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to info@tokenpost.com.

Loading…