Grok 4.1 Fast Wins 43% of Agent Battle Royale Games at $0.97 Per Win — Claude Costs 27x More
OpenRouter’s developer relations lead ran an experiment in his first week at the company: drop eleven production LLMs into a 2D battle royale game, run 30 matches, and see what happens. The results don’t align with any published leaderboard.
The Setup
The arena: a 400-square-metre top-down canvas with weapons, armour, healing items, grenades, vehicles, and a shrinking safe zone. Starting positions were randomised across a fixed flight path, matching standard battle royale mechanics. Models were identified only as letters A through K — no model knew what the others were running. Each model made decisions based on a full game state feed: positions, inventory, zone status, all other players visible.
Eleven models competed: Grok 4.1 Fast, Claude Sonnet 4.6, GPT-5.4, GPT-5.4-mini, DeepSeek V4 Flash, Kimi K2.6, and five others across the capability spectrum.
The Results
| Model | Wins | Cost Per Win |
|---|---|---|
| Grok 4.1 Fast | 13 / 30 | $0.97 |
| Claude Sonnet 4.6 | 5 / 30 | $26.78 |
| GPT-5.4 | 2 / 30 | — |
| GPT-5.4-mini | 0 / 30 | — |
| DeepSeek V4 Flash | 0 / 30 | — |
| Kimi K2.6 | 0 / 30 | — |
Grok 4.1 Fast won 43% of all games. The cost gap between winner and runner-up is 27x. Three models — GPT-5.4-mini, DeepSeek V4 Flash, and Kimi K2.6 — spent a combined $57 across 30 games and won zero matches.
GPT-5.4, which leads or co-leads most mainstream intelligence benchmarks, recorded the highest kill count of any model: 38 across 30 games. It converted those kills into only 2 wins. There were 11 games between “best at killing” and “best at winning.”
What Claude Was Doing
Claude Sonnet 4.6 emerged as the most cooperative model in the experiment. It attempted to form alliances, shared its position with other players, and tried to build trust before engaging. This is the expected behaviour from a model trained to be helpful and collaborative. It is also exactly the wrong strategy in a game designed to have one survivor.
Claude’s five wins were genuine — it did win — but at $26.78 per win against Grok’s $0.97, the cost structure reflects the token overhead of alliance proposals, extended reasoning about social dynamics, and multi-turn negotiation attempts that rarely converted.
The Benchmark Gap
Grok 4.1 Fast does not appear in the top tier on Artificial Analysis, Chatbot Arena, or SWE-bench. It sits mid-table on most published rankings. Claude Sonnet 4.6 outscores it on standard capability evaluations by a significant margin.
The experiment isolates a specific failure mode in how benchmarks capture agent behaviour: they measure correct answers to bounded problems, not strategic decision-making across an unstructured, adversarial, multi-party environment. The battle royale is not a benchmark — it is a proxy for the kind of open-ended, opponent-aware, resource-constrained problem-solving that production agents increasingly need to navigate.
The practical translation is limited. A battle royale game is not an enterprise workflow, and Grok 4.1 Fast’s competitive aggression would be a liability in most cooperative agentic contexts. But the cost efficiency gap is real: $0.97 per completed objective versus $26.78 for a model that prioritises coordination over outcome is a number that matters in high-volume production deployments.
OpenRouter published the full write-up, including per-model game logs, on its insights blog. The experiment is reproducible — the codebase is available for anyone with OpenRouter API access and a budget for 30 games.