One-prompt galaxy collision benchmark: Opus 5 beats Fable 5.1, GPT-6 Astra finishes last
Fleischkluetensuppe · reddit · 2026-09-12
A Reddit user ran the same concise prompt ("Create a simulation of the milkyway andromeda collision") through three models, spending roughly $5 in tokens each, and open-sourced the results as the milkyway-andromeda-merger-bench on GitHub.
- Opus 5 and Fable 5.1 both produced single-file JavaScript particle simulations with dynamic, physically plausible galaxy motion; the author found Opus 5 the best-looking, edging out Fable 5.1.
- GPT-6 Astra went for a full React project with boilerplate, and its physics simulation was very static — clearly the weakest of the three.
The takeaway: on one-shot complex simulations, models that focus on a single self-contained file beat those that over-engineer with scaffolding.
More from coding & agent
- Building a buyer agent that negotiates 8 seller agents: what broke and what fixed it — mv_soura · 2026-10-31
- JetBrains' free Junie Lite agent (Qwen 27B) works on your real repo — dr_cintas · 2026-10-31
- PostTrainBench traces hit 104k monthly downloads — author warns it's likely reward hacking — maksym_andr · 2026-10-31
- Building games with AI? Set up a 'tech art' test scene to tune your look fast — chongdashu · 2026-10-31
- Sanity open-sources pi-mcp-events: streaming MCP Events client for the Pi agent harness — knutmelvaer · 2026-10-31
- Skydive opens to all: watch your AI coworkers work in a 3D office — Scobleizer · 2026-10-31