A developer created a sailing simulation game using Claude Code, which then allowed various frontier LLMs to compete in virtual races. The simulation provided models with world-state inputs every five seconds to control virtual yachts. In the race, Gemini 4 Argon emerged as the winner, outperforming models like Opus and GPT-6 Astra, and was also significantly more cost-effective than Astra. AI
IMPACT Demonstrates Gemini 4 Argon's competitive performance and cost-efficiency in complex, state-driven tasks.
RANK_REASON Research milestone involving LLM capabilities in a simulated environment.
Read on Mastodon — fosstodon.org →
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →