Astra uses some of its no-CoT capability in practice

Astra scores significantly higher than previous models on no-CoT benchmarks, as for example shown in Neel Nanda's post last week. This raises the question of whether, and to what degree, Astra uses this no-CoT capability in practice.While user-facing outputs may be subject to training pressures to make reasoning intelligible, this may be less true for non-user-facing CoTs. Additionally, CoTs may be subject to training pressure to be token-efficient, particularly with Astra's cheapest and fastest...