OpenAI AI Updates: September 9, 2026

1. OpenAI Published a Navier-Stokes Result Produced by 10,000 Coordinated Agents

OpenAI. The company posted what it describes as an AI-generated solution to the Navier-Stokes Millennium Prize Problem, including a formal proof in Lean. According to reporting on the run, the result came from roughly 10,000 coordinated agents working for 88 hours on an unreleased internal model that OpenAI characterizes as significantly more capable than GPT-6 Astra, at a compute cost in the millions of dollars. The Lean formalization is the load-bearing detail for practitioners: it moves the claim from “a model wrote a plausible argument” to “a proof assistant checked it.” The announcement arrived alongside a public dispute over authorship and provenance, covered separately in today’s general roundup. Source

2. ChatGPT Images 2.5 Targets Sketches and Reference Photos

OpenAI. A new image generation release turns sketches, reference photos, and written concepts into finished images, with the pitch centered on fidelity to a user-supplied reference rather than on raw prompt-to-image quality. That framing puts it in direct competition with the reference-conditioned editing workflows that have become the practical use case for image models in production, where the constraint is usually “match this existing asset,” not “invent something new.” Source

3. An MIT Researcher Is Running Quantum Computing Experiments Through Codex

OpenAI. OpenAI documented an MIT researcher using GPT-5.6 Sol with Codex to run quantum computing experiments autonomously, including analyzing results and calibrating qubits. Qubit calibration is a closed-loop task against real hardware rather than a text task, which makes it a more interesting data point than the usual coding-agent case study: the agent’s output is measured against physical apparatus, not against a test suite it could have gamed. Source

4. OpenAI Committed $5 Million to Independent Research on AI and Teen Development

OpenAI. The grant program funds outside researchers studying how generative AI affects teen development, wellbeing, and safety. The money is explicitly for independent work, which matters because most published evidence on adolescent chatbot use currently comes from the labs whose products are under study. Source

5. 1Password Reports a 21% Engineering Productivity Gain From Codex

OpenAI. In a customer write-up, 1Password says its engineers used Codex to ship features and internal tooling faster while holding to the review requirements a password manager needs. The 21% figure is the company’s own measurement and the post does not describe the methodology, so treat it as a directional claim rather than a benchmark. Source

6. OpenAI Broadened Its Journalism Program From Newsrooms Into Classrooms

OpenAI. The company expanded its journalism initiatives to cover students, educators, journalists, and news organizations with tools, training, and partnerships. It follows the Ukraine newsroom program announced with WAN-IFRA and AIRPPU the previous day, and lands while OpenAI is defending copyright suits from multiple US publishers. Source