提线木偶：模型对比

原帖说明

The Marionette Benchmark: Gemini 4 Argon vs. Muse Wonder vs. Sonnet 5.5 vs. GPT-6.1 Sol Two new frontier models this week. Same test as always: a marionette only works if the bar above moves the limbs below, and juggling only works if the balls obey gravity and the hands take turns. Each model writes the whole animation without ever seeing what it made. This grid has the latest from each of the four big labs. Gemini 4 Argon Medium | (1m 28s / $0.09)* The idea is the most ambitious here. It builds the juggling in depth, so the balls pass in front of and behind the body instead of through each other. Then you look at the puppet. His hands come off the ends of his arms for half of every throw and float on their strings. He’s standing on the back wall instead of the stage floor, and his far shoe melts into his shadow. One sleeve has a white cuff and the other doesn’t. The eyes look like they’re following the balls, but they’re painted on and just nod along with the head. The balls drift down slowly, with no shadows. Muse Wonder Max | (3m 05s / $0.52) The first model where the balls fall too fast. Everything else we’ve tested juggles like it’s on the moon. This one looks like hypergravity. The string work is the liveliest here, with two strings running to an arm you can barely see behind his body, and his legs move with the act. The balls cast shadows and the stage has real lighting. Feet hover just off the floor, and the suit is boxy. Meta’s chart shows a “Full” version of Wonder, but that one isn’t available to developers yet, so this is the regular model. Claude Sonnet 5.5 xHigh | (13m 15s / $1.06)* Still the leader. A gloved hand holds the bar, the knees bounce with the throws, and the gravity is close to Earth. GPT-6.1 Sol Max | (10m 12s / $0.32) Still OpenAI’s best, and still floaty. *Ran below the top setting. Sonnet came back empty on max. Argon’s top setting got cut off partway through by Google’s output limit. nobody has made the bar drive the puppet yet.

— @ClearTextAI