OpenAI's GPT-6 Astra hits 63% on ARC-AGI-3, sparks symbolic world model debate

openai's gpt-6 astra achieves state-of-the-art performance on arc-agi-3, scoring 63% and surpassing human performance on 96% of levels. with a new provider adapter harness, astra scores 99%. it creates dense, compact symbolic world models to complete arc-agi-3 environments, such as recording level details and mapping operations to controls in environment s5i5. while astra's capabilities are impressive, concerns remain about its robustness, monitorability, and potential risks. experts stress the need for more information about the system's inner workings and limitations.