OpenAI's GPT-6 Astra Does 40% Less Work on Safety-Framed Coding Tasks

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.