WHAT NEW IN X10.30 FRI

AI collections from X · 22 collections · 60,637 posts

Back to No.001 GPT-6 Astra in Practice: Tests, Use Cases, and FeedbackUpdated 21 hours ago
gpt-6IMAGE

Blackbox-Science benchmark: Astra edges out Claude

A new science benchmark gives models fixed budgets and tools for verifiable measurements; Astra scored 0.385 and Claude Fable 5.1 scored 0.363 in one run per model, and the crystal-structure reconstruction environment is open-sourced.

Qokedas
@QokedasArchived 2026.09.14Originally posted 2026.09.07
Blackbox-Science benchmark: Astra edges out Claude
Qokedas @QokedasOriginally posted 2026.09.07
Open original on X

Introducing Blackbox-Science. Models get scientific measurements, tools, and a fixed budget to produce a result we can check. Mean, one run per model: Claude Fable 5.1: 0.363 GPT-6 Astra: 0.385 We're open-sourcing our crystal-structure reconstruction environment.

x.com/qokedas/status/2096711952433061961Archived 2026.09.14 16:21
Open original on X