Same-prompt test: Opus 5.5 takes 39 minutes but crushes GPT's Sol and Astra on Three.js scene
digitalml · reddit · 2026-09-30
A Reddit user ran the same prompt — build a cinematic rocket launch scene in Three.js — on GPT-6.1-Sol, GPT-Astra, and Claude Opus 5.5, all at medium reasoning:
- GPT-6.1-Sol (8m42s, 530K tokens): decent ship, weak environment, couldn't pan down to see smoke, sound bug looped forever
- GPT-Astra (9m37s, 660K tokens): better environment and UI, but an inverted cone; not overwhelmingly better than Sol
- Claude Opus 5.5 (38m54s, 6.89M tokens): nearly 5x slower but far exceeded expectations, nailing the "cinematic" ask
The author notes the gap could be closed with extra prompting, but admits he wouldn't have thought of the right instructions. He's considering dropping GPT to $20 and going Claude $200 — or a $100/$100 split due to Claude's 5-hour limit — and will run more tests at higher tiers.
More from coding & agent
- Matthew Berman is building a Dr. Mario clone for ModRetro with AI — MatthewBerman · 2026-09-30
- BYO AI subscription model falters as users grow paranoid about token consumption — perilli · 2026-09-30
- 7,042 frames, zero After Effects: recreating LOTR's map with just assets and code — Ror_Fly · 2026-09-30
- Redditor uses Claude Opus 5.5 to puppet Codex and reach GPT 6.1 Sol — Flying_Scorpion · 2026-09-30
- Swarms ships v15 Akira: 7k-star full-stack multi-agent infrastructure platform — KyeGomezB · 2026-09-30
- Hindsight Remembers, Regex Enforces: Turning Agent Lessons Into Hard Rules — Harshvithcilvari · 2026-09-30