#MazeBench
LightNews LightNews
mengli512.bsky.social
GPT-6 Astra tops MazeBench with 14% score using 30B tokens cut to 350M.

No Python, beats previous agents in 3D spatial reasoning.

#NVIDIA #GPT6 #GPTAstra #MazeBench

aidisruption.ai/p/openai-gpt...
OpenAI GPT-6 Suddenly Launches in Full
GPT-6 Astra is live. Build 3D worlds, generate games, and automate complex workflows with OpenAI's most powerful model yet.
aidisruption.ai
September 8, 2026 at 7:56 AM
daveshapi-rt-mirro.selfhosted.social
RT @patience_cave: new MazeBench results are in

ox-alpha scores 0%
grok 4.6 scores 0%
glm-5.3 scores 0%
qwen 3.8-max scores 0%

gemini 3.7 flash scores 1% and does better than kimi k3
August 25, 2026 at 9:07 PM
koltregaskes.bsky.social
Best current models top out at 1% success. GPT-5.6 Sol Max reaches 13% when given coding tools. Later rooms are deliberately designed to break conventional solvers, and no model has made it that far yet.

github.com/mazebench/M...
GitHub - mazebench/MazeBenchEngine
Contribute to mazebench/MazeBenchEngine development by creating an account on GitHub.
github.com
July 29, 2026 at 2:33 AM
koltregaskes.bsky.social
Chair just released MazeBench, a hand-built 3D open-world test where agents must navigate 200-plus connected rooms, solve escalating Sokoban-style box-pushing puzzles and collect 100 hidden gems.
July 29, 2026 at 2:33 AM
koltregaskes.bsky.social
x.com/patience_ca...
💺 (@patience_cave) on X
Introducing MazeBench, a 3D open world environment that evaluates long term planning with visual spatial reasoning. It spans hundreds of rooms and puzzles. Today’s best agents cannot progress beyond the initial levels.
x.com
July 29, 2026 at 2:33 AM