OpenAI Launches GPT-6 Astra on Sept 3 With Computer Use Feature; AI Safety Experts Warn Over Unreadable Reasoning

According to Fortune, OpenAI launched GPT-6 Astra on September 3, introducing a Computer Use feature that allows AI to directly operate computers like humans. Co-founder and President Greg Brockman demonstrated the capability in video showcases, where users issue voice commands and the AI executes tasks such as filling forms, processing spreadsheets, and browsing websites. In benchmark tests, Astra scored 98.6% on the ARC-AGI-3 reasoning test, compared to 7.8% for GPT-5.6 Sol, and achieved a perfect score on the ExploitBench cybersecurity challenge.

However, AI safety experts raised concerns about Astra's new architecture, which incorporates recurrent Transformer technology to reduce computational overhead. This design may render the model's reasoning process increasingly difficult for humans to read and monitor, potentially undermining a key safety oversight method. Former OpenAI safety researcher Stephen Adler warned that this approach could cross a critical safety line for the AI industry, while former governance researcher Daniel Kokotajlo called for industry-wide standards to preserve reasoning transparency as AI systems advance.