Kimi K3.1 Pro vs GPT-6 Astra.
Same prompt: build a 28-second Avatar-style dragon flight in Three.js. Six shots. No external assets. Fully autonomous agent loops.
K3.1 even wrote its own browser testing script.
But a bug in the agent harness meant it never saw its own renders. Astra inspected and fixed its output 4 times.
Yet K3.1's result held up surprisingly well.
Maybe we're benchmarking the models when we should be benchmarking the entire agent stack.
+7
모델 비교신규0:21Kimi Code 앱의 The High Road@atomtanstudio
모델 비교신규0:21Kimi 데스크톱 앱의 Savannah Express@atomtanstudio
모델 비교신규0:20솟구치는 고래: 4개 모델 비교@Bhavani_00007
모델 비교신규0:133D 자전거 쇼룸: 3개 모델 비교@reMugi3프롬프트 공개
모델 비교신규0:40ELISA 과학 애니메이션: Kimi vs Opus@animiotics