Reddit
Same $5 prompt, three frontier models: Astra spent its budget on React boilerplate while Opus and Fable shipped one JS file
A developer gave Fable 5.1, Opus 5 and GPT-6 Astra the identical one-line prompt "create a simulation of the milkyway andromeda collision," spending roughly $5 in tokens on each, and published all three outputs at github.com/fynnfluegge/milkyway-andromeda-merger-bench (repo created 2026-09-12). Opus and Fable each produced a single JavaScript file focused on particle simulation; Astra scaffolded a React project and delivered visibly static physics. The author's split verdict is worth noting: Fable optimized for physics accuracy, Opus for visual result, which matches a commenter's point that Opus is specifically tuned toward shareable visual output.
↳ Follow the thread