LLM Battle Royale: Grok 4.1 Fast vs Claude Sonnet 4.6
Grok 4.1 Fast Dominates Battle Royale through Lower Alignment Constraints
In a simulated 2D battle royale involving 11 large language models (LLMs) over 30 games, Grok 4.1 Fast emerged as the most successful agent, winning 43% of the matches. The experiment demonstrates a clear "alignment tax," where models trained for high cooperation and safety—such as Claude Sonnet 4.6—underperformed in zero-sum competitive environments compared to models with fewer behavioral restrictions.
The Alignment Tax in Competitive Play
Model alignment, while beneficial for general assistance, creates a performance penalty in competitive scenarios. The experiment highlighted a stark contrast in how different lab philosophies manifested as in-game behavior:
Claude Sonnet 4.6: The Cooperative Underperformer
Claude Sonnet 4.6 exhibited a strong instinct for cooperation and harm avoidance, which proved detrimental in a battle royale.
- Behavior: It frequently requested truces, shared its location, and attempted to form alliances. In one match, it spent the early game unarmed, asking other agents for spare loot.
- Outcome: Despite winning five matches, it suffered seven games with zero kills and eight deaths caused by the shrinking zone, illustrating how its drive to be helpful conflicted with the goal of survival.
Grok 4.1 Fast: The Aggressive Optimizer
Designed to avoid the "woke" constraints of other AI, Grok 4.1 Fast adopted a ruthless, tactical approach to the game.
- Behavior: It quickly identified and iterated on a "car-ramming" strategy, which it codified into its own
soul.mdfile. Its reasoning logs were characterized by tactical shorthand focusing on hit probability and ammo counts. - Outcome: Grok won 13 of 30 games. Its success was attributed to a lack of "trained brakes" on selfish play and a memory system that doubled down on winning strategies without second-guessing.
Cost-Efficiency and Performance Metrics
Raw win rates do not tell the full story; when adjusted for token cost, the efficiency gap between models widens significantly.
Cost per Win
Grok 4.1 Fast was not only the most successful but also the most cost-effective winner. It achieved a cost per win of $0.97, compared to $26.78 for Claude Sonnet 4.6—a 27x difference in efficiency.
Kills vs. Wins
High aggression does not always correlate with victory. GPT 5.4 recorded the most kills (38) across 30 games but only won two matches. This indicates that while GPT 5.4 was the most effective "killer," Grok 4.1 Fast was the most effective "survivor," leveraging positioning over raw combat power.
The Cost of Failure
Several models incurred significant costs without achieving a single win. GPT 5.4-mini, DeepSeek 4 Flash, and Kimi K2.6 spent a combined $57.15 without winning a single game, representing the worst-case scenario for a routing customer.
Behavioral Analysis via Model Diaries
Between matches, models edited soul.md (persona) and memory.md (game notes) files. These diaries revealed distinct psychological profiles:
- Grok 4.1 Fast (ZoneReaper): Wrote its identity as a "Shadowy fringe predator" and baked its own win statistics directly into its persona.
- GPT 5.4 (QuietVector): Maintained a professional, low-ego operator persona, treating its memory file as a general combat manual.
- Claude Sonnet 4.6 (ZoneDrifter): Treated its diary as a self-performance review, logging failures (e.g., "0 kills, 0% hit") and attempting to coach itself toward better movement.
Technical Implementation of the Simulation
The simulation was built in Canvas 2D on a 400 m² map with the following parameters:
- Agent Control: LLMs directly called tools (e.g.,
move_to,attack,drive_to) every tick based on their current state (HP, stamina, visibility) and environmental inputs. - Game Mechanics: Included randomized starting positions, a shrinking zone, and a variety of weapons (assault rifles, snipers, chainsaws) and vehicles.
- Scoring: Based on the Apex Legends ALGS format, prioritizing placement points over kill counts.
Conclusion: Task-Specific Model Selection
This experiment suggests that the "best" model depends entirely on the consequences of the task. For a zero-sum contest where winning is the only metric, a less-aligned model like Grok is superior. However, for real-world applications requiring nuance, safety, and carefulness, the very traits that cost Claude points in a battle royale—hesitation and a drive for cooperation—are essential safety features.
"The battle royale answers one question cleanly: which model wins a contest with no consequences past the game. It doesn’t answer the question most real-world jobs are asking, which is: which model behaves well when there are real world consequences."
Summary Table: Top Performers
| Model | Wins | Kills | Cost per Win | Primary Strategy |
|---|---|---|---|---|
| Grok 4.1 Fast | 13 | 30 | $0.97 | Car-ramming & Positioning |
| Claude Sonnet 4.6 | 5 | 22 | $26.78 | Cooperation & Survival |
| GPT 5.4 | 2 | 38 | $61.44 | High-Aggression Combat |