Building a truly novel game exposes the limits of frontier models
teortaxesTex · x · 2026-11-09
The author tried building an actually novel game with AI and observed each model's real limits: Opus proposes cutting scope out of abundance of caution, DeepSeek melts down in confusion, and Astra assumes games are slop and doesn't even try — though the author notes it's the only one fit to write the relevant math. Frontier models still each show clear weaknesses on non-templated creative tasks.
Related event: Frontier Models Fall Short When Asked to Build Truly Novel Games(2 posts)→
More from Models
- MathArena adds October BrokenArXiv: GPT-6.1 Sol leads but drifts down monthly, Opus 5.5 rock solid — teortaxesTex · 2026-11-09
- Users report Opus 5.5 back to normal, suspect Anthropic silently reverted the degradation — ryunuck · 2026-11-09
- Open-weights models now just 2 months behind closed frontier as Kimi K3.1 Pro lands — 0xsachi · 2026-11-09
- Model trained on 150,000 CVE exploits scores 88.4% on CyberGym with zero refusals — jedisct1 · 2026-11-09
- ChatGPT app appears to have nuked Chrome/browser tooling, leaving only computer use — GabGarrett · 2026-11-09
- Anthropic could start IPO roadshow as early as Tuesday, eyeing $1.1-1.3T valuation — kimmonismus · 2026-11-09