Hacker News
A Ten-Way Model/Harness Bake-Off on One Three.js Scene Shows a 47x Cost Spread and a 4x Time Spread for the Same Task
Published 2026-09-09, the test gave ten model/harness pairs the same prompt: a single-page Three.js sci-fi hangar with hovering drones, animated warning lights, emissive runway strips and volumetric-style fog planes. GLM 5.3 Flash on OpenCode finished cheapest at $0.085234 with 96.89% cached input; Astra 6.0 Max on Codex cost $4.038190 and took 37m29s. Qwen 3.8 27B x-high on OpenCode was fastest at 8m48s. Harness choice, not model choice, drives most of the spread, which matches the ARC Prize and Armature results from last week.
↳ Follow the thread