Autonomous browser-game model benchmark

MODEL ARCADE

Twenty-two autonomous coding outputs served from the same loopback origin and protected by Cloudflare Access. Select a model and game to inspect its unmodified final artifact. Entries marked failed are preserved as benchmark evidence, not certified playable results.

SPACE INVADERS — 10 MODELS

DeepSeek V4.1 Flash

DeepSeek via OpenRouter · high reasoning · independently browser-verified

GPT-6 Astra High

OpenAI via OpenRouter · high reasoning · independently browser-verified

Mellum2 12B-A2.5B

Local MXFP4 MoE · failed autonomous repair · score remained 0

GPT-5.6 Sol

OpenAI endpoint · 32K allowance

Claude Fable 5

Anthropic endpoint · 32K allowance

DeepSeek V4 Flash 0731

Morph BF16 · 32K allowance

Gemini 3.7 Flash

Google endpoint · 32K allowance

GLM-5.3

Z.AI FP8 endpoint

Qwen3.8 Max

Alibaba · completed after resume

Qwen3.8 27B

AkashML BF16

GALAGA — 10 MODELS

DeepSeek V4.1 Flash

DeepSeek via OpenRouter · high reasoning · entry, formation and dives browser-verified

MiMo 2.6 Flash

Xiaomi via OpenRouter · high reasoning · 19/19 browser checks passed

GPT-6 Astra High

OpenAI via OpenRouter · high reasoning · browser interaction verified

GPT-5.6 Sol

OpenAI · $0.1785

Claude Fable 5

Anthropic · browser crash self-fixed

DeepSeek V4 Flash 0731

Morph BF16 · extended natural completion

Gemini 3.7 Flash

Google · largest visual artifact

GLM-5.3

Z.AI FP8 · extensive verification

Qwen3.8 Max

Alibaba · real Chrome/end-state tests

Qwen3.8 27B

AkashML BF16 · required 128K continuation

3D DESCENT — 1 MODEL

GPT-6 Astra High — Deep Vector

Self-contained WebGL · six-axis flight · reactor escape · 83/83 independent checks

PIXEL ROGUELITE — 1 MODEL

GPT-6 Astra High — Pixel Panic

Procedural mall mayhem · elemental chains · Food-Court King · 41 verified test groups

No external game assets or runtime network dependencies. Anonymous access is blocked by Cloudflare Access.