DeepSeek-V4 Pro and Fable show lowest task correlation

zainhas · x · 2026-08-15

A comparison of DeepSeek-V4 Pro, Fable, and Sol reveals their behavioral similarities on tasks. DeepSeek-V4 Pro and Fable are the most diverse pair with a 0.39 correlation in outcomes, while Pro and Sol are the most alike at 0.54. The combination of all three models failed to solve only 4 tasks.

Original post →

More from Models

Models channel →