Astra built a React app; Opus and Fable wrote one file

A developer gave three frontier models the same one-line prompt, "create a simulation of the milkyway andromeda collision," with about $5 in tokens each, then published the results in a GitHub benchmark repo. Fable 5.1 and Opus 5 each wrote a single JavaScript file and put the money into the particle simulation. GPT-6 Astra scaffolded a full React project first.

The payoff split accordingly. Astra's physics came out visibly static, galaxies that don't merge the way the other two attempt. Scaffolding a project ate the budget that would have gone into the simulation itself.

Fable and Opus didn't agree with each other either. Fable leaned toward physics accuracy. Opus leaned toward the visual result. Same file format, same budget, different bet on what to spend it on.

A one-line prompt with a fixed token budget makes a model guess what the requester values, and that guess becomes spend before any actual simulation code runs. Astra guessed a React scaffold was worth building first. Fable and Opus skipped that step and put every dollar into the collision itself.

What to watch

Nobody's rerun this with an explicit constraint, like "single file, no framework," applied to all three models. That would show whether Astra's scaffolding choice is a default habit or just how it read this specific prompt. Until someone does, the result reads as one data point: on a tight budget, the model that skips infrastructure and writes straight to the problem gets more of the actual work done.