GLM-5.2 vs Claude Opus
TLDR Dev ● Covered by 3 sources
GLM-5.2, a new open-weights model, just got tested head-to-head against Claude Opus 4.8 by building a 3D game from scratch. Opus won on quality, but GLM-5.2 did it for a fifth of the price.
Someone finally ran the experiment everyone's been arguing about on Twitter: is GLM-5.2 actually as good as Claude Opus, or is this just open-source hype season again. The team at TLDR gave both models the exact same one-shot prompt — build a full 3D platformer in raw WebGL, no Three.js, no shortcuts — and let them work unsupervised. Opus 4.8 finished in 33 minutes and 30 seconds. GLM-5.2 took an hour and ten minutes. The gap in cost was even starker: $5.39 for GLM-5.2 versus roughly $21.92 for Opus at list pricing.
And the extra time didn't buy GLM-5.2 a better result. Its game shipped with a gray, untextured character whose head disappeared whenever the camera moved, a spike hazard that didn't kill the player, and a debug overlay still stuck on screen at the end. Opus's build was cleaner: textures loaded properly, the spike hazard worked, animations ran smoothly, and the character actually reached the flag to win. Opus's bugs were the kind you'd expect from a near-finished game — a slightly too-generous coyote-time jump window, a win trigger that fired a touch early — not broken fundamentals.
The real story here isn't raw capability, though. It's that GLM-5.2 is text-only, and that turned out to matter more than expected. Both models were told to verify their own work before stopping, and the standard move is to screenshot the result and eyeball it. Opus did exactly that, caught its leftover debug readouts, and cleaned them up. GLM-5.2 can't see images, so it wrote a script to sample pixel colors instead, confirmed grass looked green and coins looked gold, and declared victory — never realizing its character was rendering flat gray with a missing texture file. It was checking its work blind.
Benchmarks back up the overall pattern. Z.ai's own numbers put GLM-5.2 ahead of Opus on some reasoning tests like AIME 2026 and IMOAnswerBench, but Opus still wins most of the coding and agentic categories, things like SWE-bench Pro and NL2Repo. An independent ArtificialAnalysis run confirms GLM-5.2 leads the open-weights pack, beating MiniMax-M3, DeepSeek V4 Pro, and Kimi K2.6, but it's not close to unseating the top closed models yet.
Which is really the point. GLM-5.2 isn't replacing Opus as anyone's daily driver, and the team running this test says as much outright. But it's MIT-licensed, downloadable, and costs a fraction of the price on output tokens — under a fifth of what Opus charges. For teams building agentic coding tools where budget matters more than polish, or who just don't trust a closed lab to keep serving a model indefinitely, that's not nothing.
My take
I like open weights on principle, not because they always win benchmarks, but because they can't be pulled out from under you the way Fable and other closed products have been. GLM-5.2 losing this particular test doesn't bother me; a text-only model checking its own visual output by sampling pixel colors was always going to embarrass itself. The more interesting trend is that the gap between open and closed is now measured in minutes and dollars, not tiers of competence, and that gap is shrinking every quarter whether Anthropic likes it or not.
Read more about this at: TLDR Dev
