Escalating to the better model made 34 answers worse

Automated excerpt

Same tasks, both arms pinned, three repetitions each, so every task has a cheap answer and a frontier answer side by side. The full run is 800 standard tasks per model across three categories, plus the traps. Every other model did that between 1 and 9 times.

Selected automatically from source text; not independently written or fact-checked. Read the original for full context.