Kimi K3.1 Pro vs GPT-6 Astra on a Three.js dragon flight scene
Given the same 28-second Three.js flight task, Astra finished in 9m17s for about $2.63 and iterated on its renders, while K3.1 never saw its output due to a tooling bug.
Kimi K3.1 Pro built an Avatar-style dragon flight. It also wrote its own browser test without being asked. Then a tooling bug stopped it from seeing its own renders. An early test in Atomic Agent gave K3.1 and GPT-6 Astra the same job: Build a 28-second Three.js flight in one HTML file. Toruk + rider, six camera shots, waterfall, night jungle, sunrise. All geometry and textures in code. Both models worked independently until they decided they were done. Reported results: • K3.1: 12m 28s, 20 steps, ~$1.10–$1.40 • Astra: 9m 17s, 35 steps, ~$2.63 Kimi’s cost is an estimate using K3 rates; public K3.1 pricing wasn’t available for the test. Kimi wrote a 746-line scene, rendered seven test frames and checked for console errors. Here’s the catch: K3.1 requires temperature=1. The screenshot-review tool still sent 0.2, so both visual reviews failed. Astra could inspect its renders and made four rounds of fixes. Kimi finished without that feedback. Its reported miss was a two-second sunrise shot that rendered as a flat teal frame. That makes this a promising result, with a rerun needed for a fair comparison. My takeaway: the tools around an agent can materially change the result. Screenshot access sounds like a small detail until one model gets to see its mistakes and the other doesn’t.