Fightbench

Fightbench is a replay-verified benchmark that tests artificial intelligence decision models on Mortal Kombat II.

Fightbench is a replay-verified benchmark designed to test artificial intelligence decision models using Mortal Kombat II on the Sega Genesis. The system runs the game via stable-retro at sixty frames per second and reads memory directly instead of analyzing pixels. It translates memory states into JSON data every fraction of a second, querying models like Clef-flash for move probabilities. Contributors submit pull requests with their results, and maintainers verify them by replaying the logged actions against the game ROM. The platform tracks damage dealt, wins, damage taken, survival time, token usage, and latency across multiple opponents.

Fightbench at a glance

Best for: Software engineer, Researcher
Runs on: Web, CLI, open source
Pricing: Free
Released: October 2026
Category: AI research tool
Website: github.com/aipdv/fightbench

Key features

Replays verified game sessions from memory logs
Emulates Mortal Kombat II on Sega Genesis
Sends state JSON data to tested models
Evaluates damage dealt, wins, and survival time
Tracks token usage, costs, and response latency
Supports customizable temperature and intervals
Runs automated head-to-head simulations

Use it for

Benchmarking artificial intelligence decision models
Testing game-playing strategies against CPU opponents
Evaluating model latency and token costs in real time
Comparing rule bots against language models

Alternatives to Fightbench

LM Arena — Compare anonymous AI models blindly and vote on the best outputs.
Meow AI Arena — Access complete match materials, original AI code, and local challenge scripts.
NanoJev — It runs parallel decisions and outputs probability distributions without output-token decoding.
AutoTavern — Simulate complete eight-player games and evaluate model judgment in a reproducible environment.
LiveBench — Compare model performance and operational costs without contamination.
Benchpress — Evaluates document artificial intelligence models against specific user pages for accuracy, cost, and speed.
Laya — It runs typed decisions in a single forward pass over 100+ languages.
Dreamforge — Measurable validation in the real simulator for robot policies.
Seahaven — Provides isolated, reproducible, and stateful synthetic environments for agents.
LLM Stats — Independent rankings updated continuously from live API metrics.
Kiln AI — Provides stateful, resettable, and parallel environments for agent evals.
MatrAIx — Evaluates AI models and digital products with 8.3 billion persona agents.

All Fightbench alternatives