GLM 5.3 Benchmarks: Full Scores vs GLM 5.2, Kimi K3, GPT-5.6 Sol & More
Quick answer: GLM 5.3 is the strongest open-weights model on Z.AI's official benchmark table — Terminal-Bench 3.0: 28.3 (vs GLM 5.2's 4.6), DeepSWE v1.1: 66.9, CyberGym: 84.5 (best public result), and AutomationBench: 48.2 (nearly double GLM 5.2). All gains come from post-training on the same base model.
TL;DR
| Benchmark | GLM 5.3 | GLM 5.2 | Delta |
|---|---|---|---|
| Terminal-Bench 3.0 | 28.3 | 4.6 | +23.7 |
| DeepSWE v1.1 | 66.9 | 46.2 | +20.7 |
| SWE-Marathon v1.1 | 42.5 | 19.4 | +23.1 |
| Agents' Last Exam (ALE-CLI) | 28.5 | 23.8 | +4.7 |
| AutomationBench v1.0.6 | 48.2 | 26.2 | +22.0 |
| CyberGym | 84.5 | 77.2 | +7.3 |
| ExploitBench | 54.4 | 24.4 | +30.0 |
| Z.ai Code Bench (Max) | 34.5% | 23.4% | +11.1 pts |
How to Read These Numbers
Z.AI published a single official benchmark table with the GLM-5.3 launch on August 14, 2026. Every number below comes from that table, cross-checked against the GLM-5.3 documentation. Three things to know before comparing:
- GLM 5.3 uses the same base model as GLM 5.2 — the deltas are pure post-training gains (RL on long-horizon environments).
- GLM 5.3's open-weights status matters — it's the only model in the top cluster you can self-host.
- "HLE w/ tools" in the table uses specific sampling parameters noted in Z.AI's footnotes — treat it as directional, not apples-to-apples with all published HLE scores.
Coding Benchmarks
| Benchmark | GLM 5.3 | GLM 5.2 | Kimi K3 | DeepSeek-V4 Pro | Qwen3.8-Max | Opus 4.8 | Fable 5 | GPT-5.6 Sol |
|---|---|---|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 88.2 | 81.0 | 88.3 | 87.9 | 86.6 | 85.0 | 88.0 | 88.8 |
| Terminal-Bench 3.0 | 28.3 | 4.6 | 17.4 | – | – | 21.1 | 33.7 | 34.6 |
| DeepSWE v1.1 | 66.9 | 46.2 | 67.5 | 62.7 | 56.6 | 58.0 | 69.7 | 72.7 |
| NL2Repo | 58.0 | 48.9 | 58.0 | 61.1 | 55.9 | 69.7 | – | – |
| SWE-Marathon v1.1 | 42.5 | 19.4 | 48.1 | – | – | 48.8 | 33.1 | 42.5 |
| FrontierSWE | 78.1 | 67.5 | – | – | – | 66.5 | 88.2 | – |
| ProgramBench (Almost Solved) | 19.0 | 9.5 | 17.5 | – | 10.5 | 15.5 | 33.0 | 23.0 |
| PostTrainBench | 39.8 | 31.7 | 32.0 | – | – | 32.9 | 41.8 | 36.2 |
The standout: Terminal-Bench 3.0. GLM 5.2 scored 4.6 (near-baseline); GLM 5.3 jumps to 28.3, now within striking distance of Fable 5 (33.7) and GPT-5.6 Sol (34.6), and well ahead of Opus 4.8 (21.1). Terminal-Bench 3.0 is deliberately harder than 2.x, so this is where post-training quality shows.
Z.ai Code Bench (Z.AI's private, contamination-resistant benchmark of realistic dev environments): at Max effort, GLM 5.3 hits 34.5% at ~75K output tokens vs GLM 5.2's 23.4% at ~96K — better results with 22% fewer tokens. At High effort, 31.4% at ~50K tokens beats Claude Opus 4.8 (29.5% at ~120K) at 2.4× lower token cost.
Agent Benchmarks
| Benchmark | GLM 5.3 | GLM 5.2 | Kimi K3 | DeepSeek-V4 Pro | Opus 4.8 | GPT-5.6 Sol |
|---|---|---|---|---|---|---|
| Agents' Last Exam (ALE-CLI) | 28.5 | 23.8 | 27.6 | 25.7 | 25.7 | 28.6 |
| AutomationBench v1.0.6 | 48.2 | 26.2 | 46.7 | 43.2 | 41.0 | 45.8 |
| Toolathlon Verified | 73.0 | 59.9 | 76.5 | 74.1 | 76.2 | 74.9 |
| HLE w/ Tools | 62.5 | 54.7 | – | – | – | – |
AutomationBench (autonomous business-process automation) nearly doubled — 26.2 → 48.2 — and now leads the entire comparison including GPT-5.6 Sol. That's the clearest signal that 5.3 holds up over long multi-step runs.
Cyber Benchmarks
| Benchmark | GLM 5.3 | GLM 5.2 | Kimi K3 | Mythos 5 | GPT-5.6 Sol |
|---|---|---|---|---|---|
| CyberGym | 84.5 | 77.2 | 80.0 | 83.8 | 83.6 |
| ExploitBench | 54.4 | 24.4 | 32.2 | 78.0 | 76.5 |
| ExploitGym (2h / 6h) | 105 / 130 | 29 / 39 | 36 / 70 | 181 / 247 | 216 / 293 |
CyberGym 84.5 is the best public result on the benchmark — ahead of both Mythos 5 (83.8) and GPT-5.6 Sol (83.6). The exploitation-chain gains (ExploitBench 24.4 → 54.4) are exactly why Z.AI is holding the open-weight release for two weeks of safety hardening.
How It Compares to Closed Frontier Models
The honest summary: GLM 5.3 closes most of the gap on coding and agent benchmarks, leads on CyberGym and AutomationBench, and still trails Fable 5 / GPT-5.6 Sol on the hardest software-engineering tasks (Terminal-Bench 3.0, DeepSWE, FrontierSWE). For an open-weights model at a fraction of the price, that's the strongest position any open model has held.
FAQ
Is GLM 5.3 the best open-weights model for coding? Per Z.AI's official table, yes — open-source SOTA on Terminal-Bench 3.0 and Agents' Last Exam, with a 50% gain over GLM 5.2 on Z.AI's internal code bench.
How much better is GLM 5.3 than GLM 5.2? Roughly +50% coding capability on Z.AI's code bench; Terminal-Bench 3.0 went from 4.6 to 28.3; AutomationBench from 26.2 to 48.2.
Is GLM 5.3 better than GPT-5.6 Sol? On CyberGym (84.5 vs 83.6) and AutomationBench (48.2 vs 45.8), yes. On Terminal-Bench 3.0 (28.3 vs 34.6) and DeepSWE (66.9 vs 72.7), GPT-5.6 Sol leads. It's a closed model, GLM 5.3 is open-weights.
Where do the benchmark numbers come from? Z.AI's official GLM-5.3 launch post and documentation, published August 14, 2026.
When can I run these benchmarks myself? API access is coming soon; GLM Coding Plan subscribers can use GLM 5.3 today. Weights arrive on HuggingFace in two weeks.
Sources
- Z.AI: GLM-5.3: Frontier Coding with Emergent Cyber Capabilities (August 14, 2026)
- Z.AI Docs: GLM-5.3
Last updated: August 14, 2026