Tiny AI Arena showcases live battles between leading LLM agents

TL;DR – What Tiny AI Arena is and why it matters

Tiny AI Arena is an interactive web platform that runs head‑to‑head battles between several large language model (LLM) agents (e.g., Claude‑Sonnet‑5, Gemini‑3.6‑Flash, Grok‑4.6) on a simple grid‑based arena. By exposing match replays, leaderboards, and per‑round statistics, the site offers a concrete, observable benchmark for comparing LLM decision‑making in a multi‑agent, adversarial setting—something that traditional static benchmarks lack.

Core functionality of the arena

Live matches and replayable videos – Users can click any entry in the match history table to watch a frame‑by‑frame replay of the entire fight. Each match records the number of rounds, actions taken, kills per fighter, and damage statistics.
Leaderboard with Elo ratings – After each finished match, agents receive an Elo update (starting at 1000). As of the latest snapshot, Claude‑Sonnet‑5 leads with an Elo of 1063, followed closely by Claude‑Fable‑5.1 (1037) and Grok‑4.6 (1030).
Statistical breakdown – The table lists wins, win‑rate, total kills, damage dealt/taken, and average placement, giving a granular view of each model’s performance across dozens of games.
Game mechanics (from the README):

Goal: be the last fighter alive. Turns: each round every fighter takes one turn; turn order is randomized. Actions: move one cell, attack an adjacent enemy for 15–24 damage, or wait. Each action costs 1 AP. Obstacles: four random impassable cells per match. Power‑ups: gold grants +1 AP per turn. Kills: the killer gains +1 AP per turn and heals 50 HP (no overheal).

Observed performance trends

Claude‑Sonnet‑5 dominates early – With the highest Elo and a win‑rate of 43 % over 21 games, Claude‑Sonnet‑5 shows consistent aggression and effective positioning.
Gemini‑3.6‑Flash is a close second – Despite a slightly lower Elo (1029), Gemini‑3.6‑Flash achieves a comparable win‑rate (32 %) and often finishes with a low average placement (≈2.3), indicating strong late‑game survivability.
Grok‑4.6 excels in damage output – It records the highest total damage dealt (3676) and the most kills (35) across 29 matches, suggesting a more combat‑focused policy.
Lower‑ranked models (e.g., DeepSeek‑v4‑Flash, Kimi‑K2.6) struggle to secure wins, often ending with zero win‑rate, highlighting the performance gap between top‑tier LLMs and older or smaller variants.

Community reactions and insights

Gameplay appreciation – Users praised the aesthetic and background music, noting the site’s “fantastic” feel (cs1996). The simple yet strategic ruleset sparked nostalgia for classic programming games like Core War (rao‑v, simonebrunozzi).
Evaluation concerns – Some commenters questioned whether Elo on a grid battle truly measures “intelligence” (kouteiheika) and suggested that the leaderboard may be “iffy” for assessing LLM capabilities.
Desire for richer interaction – Several participants proposed extensions:
- Adding a shareable replay link for easier comparison (sleda).
- Allowing agents to communicate between rounds, perhaps with short text or emoji messages, to explore diplomatic or cooperative dynamics (vessenes).
- Implementing a programming layer where users could supply custom Lua bots, akin to Core War, to test LLM‑generated code (rao‑v).
Technical hiccups – Minor usability issues were reported, such as music skipping (josh‑wrale) and mobile scrolling problems (coryrc, orliesaurus).
Strategic observations – Users noted that higher‑tier models appear to “think ahead,” waiting for opponents to weaken each other before striking (isoprophlex), and that some models (e.g., Fable) adopt a passive waiting strategy that can still lead to victory (ThouYS).

Why this matters for AI research

Concrete multi‑agent benchmark – Traditional LLM evaluations focus on single‑turn QA or generation tasks. Tiny AI Arena provides a repeatable, observable environment where agents must plan, adapt, and survive against peers.
Emergent strategic behavior – The observed differences in aggression, waiting, and resource management hint at how model fine‑tuning influences decision‑making beyond language tasks.
Open data for analysis – The publicly available match logs (round count, actions, damage) enable researchers to perform post‑hoc analysis, train meta‑models, or develop new evaluation metrics.
Community‑driven extensions – The active discussion around communication, code‑generation bots, and genetic‑algorithm style evolution suggests a fertile ground for collaborative research and open‑source tooling.

Next steps for enthusiasts and researchers

Experiment with the API – Clone the repository (if available) and run custom matches, swapping in newer model versions as they are released.
Develop analytic scripts – Parse the match history CSV to compute deeper metrics such as average AP usage, kill‑to‑damage ratios, or positional heatmaps.
Propose rule extensions – Implement chat channels between agents or variable map sizes to test how communication influences outcomes.
Contribute to the leaderboard – Submit your own model runs and compare against the existing Elo rankings, helping to refine the benchmark.

Bottom line: Tiny AI Arena turns abstract LLM capabilities into a visual, competitive sport, offering a fresh lens on model behavior and sparking community ideas for richer multi‑agent AI evaluation.