OpenAI published a Codex science case study on June 11, 2026, describing how University of Arizona and Steward Observatory astrophysicist Chi-kwan Chan uses Codex to help derive, implement, and test new algorithms for black hole simulations.
The point is not that AI is directly discovering black holes. The more important signal is that AI is being placed inside a rigorous scientific workflow. Chan's team studies black holes observed by the Event Horizon Telescope and is moving from still images toward dynamic video of a supermassive black hole. That requires huge observational data pipelines, supercomputer workflows, and numerical models that can represent extreme physics.
The hard part is plasma near the black hole. Electrons and ions spiral rapidly around magnetic field lines, and standard simulations must track many tiny turns. That forces computers into extremely small time steps. Even top supercomputers can spend too much of their effort on those microscopic motions, limiting how realistic the larger simulation can become.
Chan uses Codex to propose candidate numerical methods, then his team inspects, understands, and tests them against known solutions. OpenAI is explicit that many candidate approaches are not correct. That is acceptable in science, where ideas are valuable only when they can be tested. Inspectability, reproducibility, and physical understanding are the boundary conditions.
This case matters because it extends agentic coding beyond software engineering into scientific computing. Codex is not replacing the researcher. It accelerates algorithm exploration, implementation, testing, and elimination of weak paths. People still define the question, interpret the physics, and decide what has actually been validated.
For AI workflows, this is a mature pattern. In high-value settings, the agent is often not a one-shot answer machine. It is an exploration partner inside a process. It can compress days of manual search into shorter iterations, but only when every step is bounded by tests and review.
AI's strongest scientific role may be in work that is complex but verifiable. When outputs can be checked by math, experiments, unit tests, or physical laws, speed has a better chance of turning into real progress instead of plausible prose.
