ChatGPT vs Claude Make a Kart Racing Game in one shot. One Built Mario Kart

I gave seven AI models the same prompt, to build a 3D racing kart game, then played every game myself. Four built theirs first. Then OpenAI and Anthropic released new models on the same day, 22 September 2026, and GPT-6 Sol, GPT-6 Luna and Opus 5.5 got the same prompt too. One run per game, with nothing fixed afterwards. The ranking at the end is my personal opinion. Play them yourself, free in your browser: Sunspur Kart (GPT-6 Astra): https://devarf.dev/play/sunspur-kart/ Turbo Drift Isle (GPT-5.6 Sol): https://devarf.dev/play/turbo-drift-isle/ Aster Circuit (GPT-6 Sol): https://devarf.dev/play/aster-circuit/ Cinder Circuit (GPT-6 Luna): https://devarf.dev/play/cinder-circuit/ Vroomvale (Fable 5.1): https://devarf.dev/play/vroomvale/ Sparkplug GP (Opus 5): https://devarf.dev/play/sparkplug-gp/ Pinwheel Prix (Opus 5.5): https://devarf.dev/play/pinwheel-prix/ More games built with AI: https://devarf.dev Chapters: 0:00 Seven models, one kart prompt 0:35 GPT-6 Astra: Sunspur Kart 1:15 GPT-5.6 Sol: Turbo Drift Isle 1:48 GPT-6 Sol: Aster Circuit 2:25 GPT-6 Luna: Cinder Circuit 4:03 The Codex numbers 4:37 Fable 5.1: Vroomvale 5:21 Opus 5: Sparkplug GP 6:19 Opus 5.5: Pinwheel Prix 8:04 The Claude Code numbers 8:35 Did newer mean better? My ranking How the games were made: Same user brief, default product settings, no custom skills, no added tools, no retries. The prompt's first line was "Create a new folder, then build a 3D racing kart game. Do your own research." Under it, all seven got the same list of rules. Each model built its game in its own company's desktop app, Codex or Claude Code, so each game shows what an app and a model made together, not the model alone. What each build ran on: the model as its log records it, the app version, and the day it was built. Every log records max effort. GPT-6 Astra: gpt-6-astra, Codex 0.153.4, 7 Sep 2026 GPT-5.6 Sol: gpt-5.6-sol, Codex 0.153.4, 8 Sep 2026 GPT-6 Sol: gpt-6-sol, Codex 0.155.0-alpha.16, 23 Sep 2026 GPT-6 Luna: gpt-6-luna, Codex 0.155.0-alpha.16, 23 Sep 2026 Fable 5.1: claude-fable-5-1, Claude Code 2.1.266, 11 Sep 2026 Opus 5: claude-opus-5, Claude Code 2.1.266, 11 Sep 2026 Opus 5.5: claude-opus-5-5, Claude Code 2.1.280, 23 Sep 2026 Sol's two builds, and Opus's two, also ran on different app versions and days, so what changed between them isn't down to the model version alone. About the numbers: Working time is how long each model spent working, not counting any time it sat blocked. Prices are what the tokens would cost at list price, not money anyone paid. I use subscription plans. List prices are OpenAI's and Anthropic's published API prices, read 22 to 26 September 2026. The Codex prices use the standard rate for shorter requests, so they are a minimum. One run per game is one example, not a benchmark. Where I say two games look alike, or guess why, that's my own impression. The logs show that no build opened another game's folder. What a model says it tested comes from its own report.