A 30-game battle royale was conducted across eleven large language models (LLMs) to evaluate their performance, costing $482 in inference.

In this experiment, Grok 4.1 Fast emerged as the top performer, winning 43% of the matches, while Claude Sonnet 4.6 consistently attempted to collaborate with others. Interestingly, the least expensive model outperformed the priciest by 27 times in terms of cost per win. The findings highlight a discrepancy between typical benchmark results and practical applications of these models.

Source: openrouter.ai