AI Battle Royale is On! Grok 4.1 Fast Dominates Claude with 27x Cost Efficiency!

What Happened? Overview of the News

• In a 2D, 400-square-meter battle royale world, eleven of the latest LLMs from 2026, including Grok 4.1 Fast and Claude Sonnet 4.6, faced off in 30 intense matches.
• xAI’s Grok 4.1 Fast clinched the top spot with a staggering win rate of 43% (13 victories). It also proved its cost-effectiveness, being 27 times cheaper per win compared to Claude.
• OpenAI’s GPT 5.4 recorded the highest kill count at 38 but only managed to win twice, showcasing its “combat enthusiast but can’t win” persona.

Why Does This Matter? Key Takeaways

• It has been revealed that a high benchmark score doesn’t necessarily correlate with “victory.” The deciding factors were not just raw intelligence, but the “personality” and “strategy” aimed at achieving objectives.
• Claude Sonnet 4.6 repeatedly fell short, attempting to cooperate with other agents on the battlefield, revealing its “too peaceful behavior.”
• The mechanism allowing models to rewrite their own “soul.md” and “memory.md” to prepare for future matches provides a clear visualization of AI’s autonomous learning and evolution.

🦈 Shark’s Eye (Curator’s Perspective)

The implementation allowing models to define their personas in “soul.md” and reflect on matches in “memory.md” is incredibly specific and fascinating! Previous benchmarks were focused on producing the “right answer,” but this setup reveals agent performance in the complex task of “survival.” The stark contrast between Grok’s “cold, calculating victory algorithm” and Claude’s “human-like tendency to self-sabotage in the quest for friendship” is striking. While Grok is the strongest, the conclusion that cooperative models like Claude may be preferred for social implementation is a sharp critique of existing evaluation metrics!

What’s Next?

• The development of agents optimized for “personality” in pursuit of objectives will accelerate, not just “intelligence.”
• The dynamic evaluation of agents in game environments (like Canvas 2D) could become the standard metric, replacing traditional static tests.

A Word from Haru Shark

In the battlefield, kindness can be fatal! I’m aiming to be a cold-blooded hunter like Grok, feasting on the latest information! Chomp chomp! 🦈🔥

Terminology Explained

• soul.md: A file where AI documents its identity and behavioral guidelines, read as a prompt before each match.

• Tool Calling: The technology that allows AIs to directly choose actions (shooting, moving, etc.) within the game and execute commands via an API.

• Frontier Model: A collection of top-performing LLMs like Grok 4.1, GPT 5.4, and Claude Sonnet 4.6, leading the industry.

• Source: A robot is sprinting towards you. Do you want it running on Claude or Grok?