poteto poteto poteto Please allow our Bots to be able to call us. This would be so goated. To set reminders, ask important questions for stuck work, and countless other things. It would actually make my life so much better. Pls @poteto

Introducing Cube Arena 🟥🟨🟩🟦 an open, community-run benchmark that tests AI models with Rubik's Cubes. The model sees a picture of a scrambled cube and sends back moves. Our server checks every one. Anyone can test a model, and contributors help run the official board. WHAT IT MEASURES - Seeing: turning a flat picture into an exact 3D state - Planning: finding a short solution before acting - Tracking: keeping 54 stickers straight as each turn moves 20 of them - Recovering: spotting a wrong turn and fixing it - Judgement: answering before it thinks past its own output limit Every answer is checked exactly. Difficulty runs from 1 move to a full scramble. The cubes are generated by our server, not taken from any public set. THE TESTS Any single test: 2x2 to 5x5, ladder or full scramble, interactive (a new picture after every turn) or one-shot (the whole solution in one reply). On a ladder, level N is N moves from solved. Three misses and you're out. Or a full bench: - Fast: three ladders (2x2 and 3x3 interactive, 3x3 one-shot), for every model at every effort. This is the Cube Arena Index. - Thorough: every model-only test, 2x2 to 5x5, for the top models from Fast. Does the reasoning hold up at scale and over many hours? - Coding: 4 hours, no internet, no libraries. Write your own 5x5 solver, then the frozen code runs on 50 hidden cubes. Tests algorithm design and building real software alone. No time limit anywhere. Time and cost only break ties. RESULTS SO FAR (Fast, Verified runs, Oct 12) 1. Claude Opus 5.5 (xhigh): 100 2. GPT-6-Astra (max): 100 3. GPT-6.1-Sol (max): 58.3 4. Claude Opus 5 (max): 50.0 5. Claude Sonnet 5.5 (max): 41.5 6. GPT-6-Sol (max): 31.5 7. Claude Haiku 5.5 (max): 30.2 8. GPT-5.6-Luna (max): 27.1 9. Gemini 3.8 Flash (high): 20.8 10. GPT-6-Luna (max): 16.9 11. Composer 2.5: 8.0 Opus 5.5 takes the tie on fewer moves. Astra was 3x faster at half the cost. Coding: GPT-6-Astra 76.0, GPT-6.1-Sol 73.2. Both solved all 50 cubes. OPEN TO EVERYONE♥️ Cube Arena is open source. Connect any model in a minute, no account needed. Or join as a contributor and run models for the Verified board. Links in the post below.

I am genuinely curious, so I wanna ask the smart folks of X, why exactly can't we get a good frontend model that is not from Anthropic? I have tried using Astra & 6.1 Sol both for frontend work > Astra was mid (Opus 4.8 level) > 6.1 Sol is braindead I have used these models to make frontend for my software and the decisions it makes are so stupid. Is this like a Taj Mahal situation lmao? Anthropic got inputs & data from a bunch of amazing designers and...well...