An astrophysicist uses Codex to simulate black holes

Simulating a black hole is not a user interface problem: it's a computational physics problem, dense with partial differential equations, numerical discretization, and industrial quantities of specialized code that few people in the world know how to write. Chi-kwan Chan, an astrophysicist at the University of Arizona, is one of those people. And according to his account on the OpenAI official blog, he's been using Codex for months to accelerate precisely that work.

The article, published on June 11, 2026, is not a generic case study. Chan describes concretely how he delegates to Codex the writing of code snippets for his relativistic plasma simulations around compact objects, the type of physical environment that the Event Horizon Telescope first photographed in 2019. The goal is not for Codex to "understand" the physics; it's to free Chan from repetitive implementation tasks so he can focus on model design and results interpretation.

What Codex does here that an IDE doesn't

The distinction matters. Codex doesn't act as a glorified autocomplete: Chan uses it in agent mode, with persistent context about the simulation codebase, to generate complete routines, refactor modules, and debug errors in high-performance code (usually C++ or Python with numerical extensions). The difference from a conventional code assistant is that Codex can maintain coherence across files and reason about project architecture, not just the current line.

This fits the pattern we've been seeing for months in tools like Claude Code: the real value of code agents isn't in writing trivial functions, but in sustaining enough context to work on projects of some scale without the researcher having to repeat the state of the world at each conversation turn.

Why it matters in computational science

Theoretical physics and computational astrophysics have a well-documented bottleneck problem: researchers with enough domain knowledge to design a simulation aren't always fluent software engineers, and engineers who could implement it efficiently don't have the physics background to know if the result makes sense. Tools like Codex don't completely solve that gap, but they narrow it.

In Chan's specific case, the simulations serve to test predictions of Einstein's general relativity in extreme regimes—intense gravitational fields, relativistic speeds, spacetime dragging effects—where no physical laboratory is possible. Every simulation cycle that gets cheaper or faster is, in practical terms, more hypotheses tested within the same time budget.

The model used matters (and OpenAI doesn't fully specify it)

The OpenAI article doesn't detail which exact version of Codex Chan is using or under what infrastructure it runs. OpenAI relaunched Codex in May 2025 as a cloud-based coding agent integrated into ChatGPT, different from the original 2021 model. We assume it refers to that more recent iteration, but it's worth keeping in mind: "Codex" today is a product with its own identity, not just a model name.

For teams working with Claude Code instead of the OpenAI ecosystem, the direct parallel would be combining specialized sub-agents with an MCP server that exposes the simulation repository as structured context. The usage pattern is transferable; the implementation is not.

Who should read this

Mostly three profiles: researchers in computational sciences who haven't yet integrated code agents into their workflow and are looking for evidence of real usage (not demos); engineers building integrations for scientific environments and wanting to see what kind of context an agent needs to be useful in that domain; and research group leaders making decisions about which AI infrastructure to adopt.

What Chan's case illustrates clearly is that the value of these agents in science isn't replacing the expert, but reducing friction between what the expert knows and what the code does. That's a considerably more modest goal than "solving physics," and precisely for that reason it rings true.

---

Editor's note: cases like this are what interest us most at ClaudeWave, not because they validate any narrative about artificial general intelligence, but because they show where the concrete benefit is today: in reducing specific bottlenecks for people with deep knowledge and scarce time. More of this, less demos.