GLM 5.2 vs. Opus
Points and comments are a snapshot, not live.
A head-to-head test finds Claude Opus faster and cleaner, GLM-5.2 far cheaper.
The article compares GLM-5.2 and Claude Opus 4.8 by having both build a 3D platformer in raw WebGL from a single prompt. Opus finished in 33 minutes, produced a cleaner game with proper textures and working win condition, and cost about $22. GLM-5.2 took 71 minutes, had missing textures and broken hazards, but cost only $5.39. The author notes GLM-5.2 is text-only, limiting its self-verification on visual tasks, while Opus used its multimodal ability to inspect a screenshot and fix issues. Benchmarks show GLM-5.2 close to top closed models on several coding and reasoning tests, though behind on SWE-bench Pro and NL2Repo.
The author concludes they are not switching their main model from Opus, but will keep GLM-5.2 as a cheaper, open-weight alternative that cannot be revoked.
What commenters are saying
Commenters generally agree the single one-shot test is not a rigorous benchmark, though they acknowledge its value for testing long-running agentic tasks. Several users share hands-on experiences: GLM-5.2 produces good output but is slow and inefficient with tokens, often straying during planning. One user notes it burned nearly $10 in an hour of work, making it less cost-competitive than API rates suggest. Another appreciates that GLM-5.2 reveals its full reasoning trace, allowing early correction. Two camps emerge: those valuing GLM-5.2's open weights, transparency, and low self-hosted cost, and those finding it too slow and token-hungry versus GPT-5.5 or Opus for practical daily use.
Several comments also note that subscription pricing (e.g., Claude Max vs. z.AI plans) may shift the cost comparison significantly.