GLM-4.5 (Thinking): benchmark results
GLM-4.5 evaluated with thinking enabled. Provider: Zhipu. Released 2025-07-28. Access: Open.
Aggregate Rating 1617 ± 20, rank #657 of 2168 rated models, from 28 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| BenchTable - Utility | 77.7 | Weighted Score (%) | 93.8 |
| LLM2014 Code 2025-09 - TypeScript | 8.89 | Score | 92.1 |
| AgentNoiseBench (Search, Tool Noise) | 37 | Accuracy on 500 sampled 2WikiMultiHopQA and HotpotQA questio | 91.3 |
| LLM2014 Code 2025-09 - C++ | 5.62 | Score | 84.2 |
| LLM2014 Code 2025-09 - Python | 7.15 | Score | 78.9 |
| BenchTable - Reasoning | 62.4 | Weighted Score (%) | 78.8 |
| AgentNoiseBench (tau2-bench, Tool Noise) | 49 | Success on a stratified 25 percent sample of tau2-bench task | 78.3 |
| BenchTable | 60.8 | Total Score (%) | 71.7 |
| AgentNoiseBench (VitaBench, Tool Noise) | 35 | Success on a stratified 25 percent sample of VitaBench tasks | 69.6 |
| AgentNoiseBench (VitaBench, User Noise) | 37 | Success on a stratified 25 percent sample of VitaBench tasks | 69.6 |
| AgentNoiseBench (tau2-bench, User Noise) | 63 | Success on a stratified 25 percent sample of tau2-bench task | 69.6 |
| MedMT-Bench (Text Subset) | 40.85 | Accuracy (%) on the 285 text-only dialogues of MedMT-Bench's | 66.7 |
Interactive version: theaggregate.ai/model?slug=glm-4-5-thinking · How It Works · Data refreshed daily, snapshot 2026-10-20.