Z.ai released GLM-5.2, an open‑weight model aimed at long‑horizon coding tasks. The Hacker News community immediately benchmarked it against Claude Opus 4.8, feeding both the same one‑shot prompt to build a full 3D platformer in raw Web GL with no external engine. Identical Kenney CC0 assets ensured a fair comparison, and the runs were timed and costed.

Opus completed the engine in 33 minutes, producing a polished game that handled camera control, collision detection, and win conditions correctly. Its Claude Code interface logged roughly $5.39 in real‑billed usage. GLM‑5.2 required 1 hour 10 minutes, generated more output tokens, and under list pricing cost about one‑fifth of Opus, confirming the headline cost advantage.

A key distinction emerged in self‑validation: Opus captured a screenshot of its output, inspected visual elements, and fixed missing hazards. GLM‑5.2, being text‑only, resorted to reading raw pixel values, misinterpreting colors and declaring the game complete despite absent win logic and texture glitches. This multimodal gap directly affected final quality.

The experiment underscores a trade‑off: Opus offers speed, visual verification, and higher fidelity at a premium, while GLM‑5.2 delivers comparable coding stamina for a fraction of the price and guarantees perpetual access through MIT‑licensed weights. Teams can now pick a closed service for production polish or an open model for budget‑constrained, self‑hosted pipelines.