GLM-5.2 vs. Opus 4.8: cheap open weights, but Opus still ships the better game

Z.ai’s new GLM-5.2 is an open-weights (MIT-licensed) flagship aimed at long-horizon agentic coding, with a 1M-token context window and High/Max thinking tiers. To cut through the launch hype, the author ran it head-to-head against Claude Opus 4.8 on a single demanding prompt: build a 3D platformer in raw WebGL with no game engine or library. The task was chosen because it can’t be faked in one pretty file — it forces a GLB parser, matrix and quaternion math, GLSL skinning shaders, substepped collision, and a follow camera, testing both multi-step agentic stamina and the reasoning needed to get engine internals right.

Opus won the practical comparison. It finished in roughly half the wall-clock time and produced a cleaner, working game with proper textures, smooth animations, a functioning spike hazard, and a real win condition. GLM-5.2 ground on far longer but shipped a rougher result: missing character materials, a spike that doesn’t kill, and a flag that triggers no win. A key differentiator was self-verification — Opus is multimodal and screenshotted its own output to catch defects, while text-only GLM-5.2 couldn’t see its render and fell back to scripting raw pixel-color sampling. Because the expected colors were present, it declared success and missed the visibly broken textures and a leftover debug overlay.

The verdict is nuanced rather than dismissive. GLM-5.2 costs a fraction of Opus — under a fifth the price on output tokens — and runs locally via vLLM, SGLang, or Transformers, so it earns a permanent place in the toolkit. The author also makes a durability argument for open weights: a closed model can be retired or restricted with little notice (citing Fable as a recent example), whereas downloadable weights can’t be taken away. Still, for workflows that depend on visual self-checking or top-tier correctness, Opus remains the default.