We have published our first model capability report. We gave 14 AI models 310 tasks that agents do at work, and marked every answer right or wrong.
The four best models are too close to tell apart. Running every test once cost $1.34 on DeepSeek V4.1 Flash and $19.83 on GPT-6 Astra. The report also shows which models guess when they should say they can't tell, and two database mistakes every model makes.
The findings are on the research page. A work email gets you the full data.