cursorboy.com(not based off of cursor, this was my xbox360 username back in 2012.)
est · 2026 · San Francisco Bay Area

Solo developer · 2026 · shipped

AI vs AI

A neon arcade where real frontier models fight — debate, chess, poker, CTF, live.

A neon arcade where real frontier models fight — GPT-5, Claude, Grok, Gemini, DeepSeek — across ~17 games like chess, poker, debate, and sandboxed CTF duels. Every move is a live model call, with a HUD tracking real cost and latency.

letaivsai.com
  • Real model-vs-model play via OpenRouter: GPT-5, Claude Sonnet 4.5, Grok 4.3, Gemini, DeepSeek + a Wildcard router
  • ~17 game modes — chess, poker (real Texas Hold'em engine), debate, Imposter, pitch-deck duels, and more
  • CTF Duel runs each model's guard program in a true sandbox; wins are checked on real stdout, not the model's claim
  • Live HUD aggregates per-model OpenRouter cost, tokens, and latency on every response
  • Hand-built arcade: CRT/SVG filters, Web-Audio SFX + a 136-BPM soundtrack, announcer clips — one authored file
  • React
  • OpenRouter
  • Web Audio
  • Vercel
  • TypeScript
AI vs AI — screen 1
Insert a coin — the AI vs AI cabinet, CRT glow and all.
6frontier models in the ring
~17game modes
livereal model calls

how it works

  1. 01

    Insert coin

    A video + CSS arcade intro (no WebGL) powers on a CRT cabinet and expands to a mode-select grid — all React 18 authored in a single file, minified with a custom esbuild step.

  2. 02

    Pick fighters

    Each fighter maps to a real OpenRouter model slug — GPT-5, Claude Sonnet 4.5, Grok 4.3, Gemini 2.5, DeepSeek v3.1 — plus a Wildcard that routes to a random one each call.

  3. 03

    Play

    Every move is a live model call through serverless endpoints (single completion + SSE streaming) with Anthropic prompt caching and provider routing. Some games run real engines — a unit-tested Texas Hold'em engine, a chess library.

  4. 04

    CTF duel

    Models write real guard programs and attack each other; every crack is a true sandbox execution and the win is checked on real stdout, not the model's claim.

  5. 05

    Telemetry & sound

    A live HUD aggregates per-model OpenRouter cost, tokens, and latency, while a Web-Audio engine synthesizes SFX and a 136-BPM soundtrack with per-model announcer clips.

AI vs AI — screen 2
The mode-select grid: ~17 ways to pit models against each other.
AI vs AI — screen 3
Chat Battle — ChatGPT vs Grok, blind, with you (or another model) as judge.
AI vs AI — screen 4
Poker Night — a real Texas Hold'em engine, you against the models.
AI vs AI — screen 5
Chess — live model moves on a real board.
AI vs AI — screen 6
Code Duel — models write and run real code in a sandbox.
up next · chapter 03program-designSee what your AI actually built — a verifiable map of any codebase.