GoldieBench deep dive
GLM-5.2's full benchmark breakdown.
The never-forgets agent — 1M context, open weights. Every number below is either one of my own judged one-shot builds - playable on this site - or an outside result with its source linked. Nothing display-only, nothing vibes.
01 · The headline numbers
02 · Every benchmark, bar by bar
All 47 scored tasks, best first. Hover for the judge's comment; click through for every model's build on that task. Or overlay other models in the interactive graphs.
03 · Where it wins - with the judge's own words
“GLM filled the bowl with glowing liquid that actually sloshes — the most convincing 'liquid in a bowl'. Opus's particles glowed but clumped to the centre. Kimi's collapsed into a tiny blob.”
“Funniest result of the lot: GLM and Opus independently produced near-identical premium 'Introducing Nova 1 — Intelligence, reimagined / distilled' keynote pages — gradient hero, full nav, pricing tiers. A dead heat. Kimi's was a plainer set of feature cards.”
“GLM's is the most cinematic — neon towers, a setting sun, Japanese signage and a flight HUD, like a frame from a film. Opus's is a clean canyon of lit skyscrapers racing to a vanishing point. Kimi leaned into the synthwave sun and grid more than the city itself. GLM wins the skyline.”
“This is GLM's. A cyan wireframe mountain range scrolling under a scanline synthwave sun — the single most beautiful frame in the whole shoot-out. Opus's clean Tron grid and magenta horizon is a close, cooler-toned second. Kimi got the idea but blew the exposure — the grid washes out to near-white. GLM wins this one going away.”
“GLM built the densest, most detailed city — windowed skyscrapers, a speed + coins HUD. Opus ran the furthest with the cleanest motion (Score 303). Kimi's runner plays fine but is unforgiving — it crashes within seconds.”
04 · Where it struggles - quoted, not hidden
Boards that hide the weak rows are brochures. These are GLM-5.2's lowest scored builds, verdicts unedited:
“Kimi nailed it — brick walls, a checkered floor, a clean minimap, textbook Wolfenstein, runs clean out of the box. Opus's is close and more atmospheric: warm fog and a vignette down a stone corridor (A/D to turn, W/S to move). GLM's engine is genuinely good — brick and mossy-stone walls, fog, a minimap — but its one-shot spawned the player buried inside a wall, dead on arrival; I nudged the start one cell so you can actually walk it. That spawn bug is why it scores lowest here, even though the e”
05 · Category breakdown vs the whole field
Games (23 tasks)
Others (0 tasks)
Pages (3 tasks)
Sims (12 tasks)
Visuals (9 tasks)
06 · Outside signals - every row sourced
My bench measures one thing: judged one-shot builds. These outside rows measure other things - kept separate, never blended into the GoldieBench average, every value linked to where it comes from.
Vendor-published numbers
| Benchmark | Result | Source |
|---|---|---|
| SWE-bench Pro | 62.1% (vs GPT-5.5's 58.6%) | groundy.com |
| FrontierSWE | 74.4 - within a point of Opus 4.8's 75.4 | venturebeat.com |
| Terminal-Bench 2.1 | 81.0 (82.7 with best harness) | groundy.com |
| AIME 2026 | 99.2 | groundy.com |
| GPQA Diamond | 91.2 | groundy.com |
From the source guides
| Measure | Result | Source guide |
|---|---|---|
| SWE-bench Pro | 62.1% | /glm-5-2-benchmarks |
| vs GLM-5.1 on SWE-bench Pro | +3.7 pts (was 58.4%) | /glm-5-2-benchmarls |
| Context window | 1M tokens | /glm-5-2-benchmarks |
| Cost vs GPT-5.5 | ~1/6 (~$1.40 in / $4.40 out) | /glm-5-2-benchmarks |
| Headline | Beats GPT-5.5 on multiple long-horizon coding benchmarks | /venturebeat-2026-06-17 |
07 · Head-to-head records
08 · Methodology + honest limits
Every GoldieBench score is a one-shot, single-file build from an identical prompt - no retries, no hand-fixing - rendered for real, screenshotted, and scored 0-10 by one judge model on one rubric across the whole field. Failures score as failures. What this bench does NOT measure: multi-turn agent work, long-context recall, or API latency - that is what the sourced outside rows are for. Full method: /methodology.
Frequently asked questions
01How does GLM-5.2 perform on GoldieBench?
GLM-5.2 averages 7.77/10 across 47 scored one-shot build tasks, ranking #10 of 17 ranked frontier models, with 5 task golds, 0 silvers and 2 bronzes. Every score is a real judged build you can open and play on this site.
02What is GLM-5.2 best at?
Its strongest scored build is Fluid at 9.0/10. By category it averages Game 7.8, Page 8.0, Sim 7.8, Visual 7.7 on the bench.
03How much does GLM-5.2 cost?
Open-weights release: weights downloadable from Hugging Face for self-hosting, or runnable for free on z.ai for individuals (commercial use has separate licensing).
Source ledger
- 01Official vendor site: z.aiz.ai
- 02SWE-bench Progroundy.com
- 03FrontierSWEventurebeat.com
- 04SWE-bench Proagentos.guide
- 05vs GLM-5.1 on SWE-bench Proagentos.guide
- 06Headlineagentos.guide
- 07Julian's guide: /glm-5-2agentos.guide
- 08Julian's guide: /glm-5-2-hermesagentos.guide
- 09Julian's guide: /glm-5-2-free-geniusagentos.guide
- 10Julian's guide: /glm-vs-kimi-vs-opusagentos.guide
- 11Julian's guide: /glm-vs-qwen-vs-opusagentos.guide
- 12Julian's guide: /three-dragonsagentos.guide
- 13Julian's guide: /the-content-machineagentos.guide
- 14Julian's guide: /the-everywhere-engineagentos.guide
Every model's breakdown
Run this stack yourself.
Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 4,000+ founders shipping with it every day all live inside the AI Profit Boardroom.