Epoch AI evaluated six language models on open-ended tasks drawn from the real work of its employees, including graphic design and research design. No model met the organization’s human quality standard, with the best performers averaging approximately 65%. Claude Fable 5.1 and GPT-6 Astra shared the top spot, while Gemini 3.8 Flash ranked last with 42%. The models performed better on coding and data analysis but struggled with tasks requiring judgment or adaptation to a particular audience.
Source: Read the original article


