GLM-5.2 vs Opus: head-to-head in a one-shot 3D game build

GLM-5.2 is Z.ai's latest flagship model: open weights under an MIT licence, a 1M-token context window and two thinking effort levels, High and Max. It is built for long-horizon coding and agent work, and it is text-only — it cannot read images. To see what that means in practice, Tech Stackups ran it head-to-head against Claude Opus 4.8 with one identical one-shot prompt: build a 3D platformer from scratch in raw WebGL, with no game engine and no 3D library.

The head-to-head numbers

Both models got the same CC0 asset pack from Kenney and one attempt, with no hints. GLM-5.2, running in Pi over OpenRouter, took 1 hour 10 minutes 40 seconds and 128 tool calls, produced 131,000 output tokens and peaked at 16 per cent of its context window; the run cost $5.39, billed for real. Opus 4.8, in Claude Code with extended thinking on high, finished in 33 minutes 30 seconds, used 153 tool calls, produced 216,809 output tokens and was estimated at about $21.92 at list prices.

What each model built

Both games are third-person platformers with the same controls: move with WASD or the arrow keys, jump with space, sprint with shift, orbit the camera with the mouse. GLM-5.2's build came out rough. Its character renders flat gray because the renderer never loaded the shared palette file the Kenney models point to, its head disappears when the camera moves, the death spike does nothing, and a debug overlay was still on screen at the end. Its spring launch mechanic, though, worked.

Opus's game was cleaner: textures and animations were applied properly, the controller and camera worked, the spike killed the player and there was a real win condition. Its bugs were edge cases — the coyote-time grace period was tuned so generously that the character could stand in mid-air beside a platform, and the win triggered while the character was still short of the flag.

Self-checking

Both models were asked to verify their work. Opus could read a screenshot of the rendered scene and removed leftover debug readouts; GLM-5.2, being text-only, sampled pixel colours with scripts and never saw that its character was gray.

Benchmarks and price

On Z.ai's own model card, GLM-5.2 leads some reasoning rows, scoring 99.2 on AIME 2026 and 91.0 on IMOAnswerBench against Opus 4.8's 95.7 and 83.5, while Opus takes most coding and agentic rows — 69.2 versus 62.1 on SWE-bench Pro and 69.7 versus 48.9 on NL2Repo. Artificial Analysis ranked GLM-5.2 the leading open-weights model on its Intelligence Index with a score of 51, though it noted the model burns about 43,000 output tokens per task, up from 26,000 for GLM-5.1.

The price gap is the headline: per million tokens, GLM-5.2 costs $1.40 in and $4.40 out, against $5 and $25 for Opus 4.8. The weights sit on Hugging Face and ModelScope under MIT, with no regional restrictions, and can be served locally with vLLM, SGLang or Transformers.

The verdict from the test: use GLM-5.2 when cost and openness matter and the work is mostly text and logic; pay for Opus when correctness, polish and visual judgement matter. And keep GLM-5.2 around regardless, because weights you can download cannot be withdrawn by a vendor.