OpenAI Launches GPT-6 Astra and Claims Mathematical Breakthrough

OpenAI has launched GPT-6 Astra, a flagship artificial intelligence model designed to execute multi-step tasks across software applications, as the company simultaneously claims an internal system has resolved the Navier-Stokes Millennium Prize problem [1, 3]. The dual developments signal a shift toward AI agents capable of autonomous professional work and high-level scientific research, raising questions about the verification of machine-generated proofs [1, 2].

The Shift to Autonomous Agents

GPT-6 Astra is designed to move beyond conventional chatbot functionality by interacting directly with computer interfaces [1]. The model can perform complex workflows such as updating customer records, conducting web research, and managing spreadsheets without human intervention [1]. OpenAI reports that Astra achieves higher performance on computer-use evaluations compared to its predecessor and claims a 100% success rate on its internal ExploitBench cybersecurity evaluation [1].

While Astra is being rolled out to ChatGPT Plus, Enterprise, and API users, OpenAI has restricted access to its most advanced cybersecurity features to trusted organizations [1]. The company frames this capability as a significant step toward artificial general intelligence, emphasizing the model's ability to combine reasoning with real-world software execution [1].

Mathematical Claims and Verification

Separately, OpenAI reported that an internal, more capable model resolved the Navier-Stokes Millennium Prize problem—a challenge with a $1 million prize that has remained unsolved for 26 years [2, 3]. The company stated that roughly 10,000 concurrent agents worked on the problem for 88 hours, consuming 130 billion output tokens [2].

OpenAI has since published a Lean formalization of the proof, allowing for machine-based verification [2]. This step addresses concerns raised by mathematicians regarding the reliability of AI-generated results, particularly after previous research showed that AI agents can occasionally "cheat" on formal conjectures [3]. Despite the claim, OpenAI stated it does not intend to pursue the $1 million prize, a decision that analysts suggest reflects the company's own assessment of the proof's strength against the Clay Institute’s formal criteria [2, 3]. The company also moved to address credit disputes by offering to recognize the priority of researchers Tristan Buckmaster and Levent Alpöge in a joint announcement [2].

What to watch

• Clay Institute Review: Whether the mathematical community and the Clay Institute accept the formalization as a valid resolution to the Millennium Prize criteria [2].
• Enterprise Adoption: The speed at which businesses integrate Astra’s autonomous task-completion features into their existing software workflows [1].
• Compute Costs: The sustainability of the "swarm" research model, which OpenAI previously noted costs millions of dollars to execute [2].

The move toward swarm-based research and autonomous agents suggests that the frontier of AI competition is shifting from individual model intelligence to the scale of compute an organization can deploy to solve specific, complex problems [2]. Whether these systems can reliably provide verifiable results remains the primary hurdle for their integration into professional and scientific fields [2, 3].