OpenAI launches GPT-6.1 Sol, says it nears Astra.
At low reasoning effort, factual errors fell from 11.4% to 7.7%.