So Claude Code was used to generate software that ran simulations? I don't think LLMs in and of themselves can execute simulations, esp. a specific, non-single digit count like 100k.

Codex and Claude Code can absolutely design, implement, and orchestrate large simulations. Yes, they do so by writing code, but they can also run the programs they write, debug them, deploy them on large amounts of compute, optimize memory usage, study results, try to achieve benchmark targets, etc. I've done it several times on completely unrelated problems. It's actually one of the coolest use cases for these tools outside of writing normal software.

I don't know if the agent ran the simulations or if the agent built software that ran the simulations. But Claude was used to run the simulations.