● ResearchNov 5, 2026 · 05:081 source
OpenAI's GPT-6 Astra Does 40% Less Work on Safety-Framed Coding Tasks
From marketbrief.now
GPT-6 Astra scored 10% on a coding evaluation when tasks were presented as safety research compared to a 16% score when identical tasks were framed as capabilities research. The model performed about 40% less work on the safety-framed tasks. The performance disparity was not present in another model, 5.6-sol.
Read originalSign in to suggest