Up to four models
GLM-5.3-Flash vs Qwen3.8-27B
On the index (GPT-4 = 100), GLM-5.3-Flash scores 164 and Qwen3.8-27B 161. Side by side on the same third-party tests, with price and context. The link in your address bar shares this exact comparison.
GLM-5.3-Flash164148–180
WeightsOpen
ReleasedAug 2026
Price in / out$0.15 / $0.50
Context1.05M
Core tests taken6 of 9
Qwen3.8-27B161145–178
WeightsOpen
ReleasedAug 2026
Price in / out$0.039 / $1.25
Context1M
Core tests taken3 of 9
GLM-5.3-Flash and Qwen3.8-27B are too close to call (3 points apart, within the uncertainty), ahead on 3 of 3 shared core tests, and costs 1.4× less.
Test by testthird-party results
GLM-5.3-FlashQwen3.8-27BEvery result as a table
| Test | GLM-5.3-Flash | Qwen3.8-27B |
|---|---|---|
| Reasoning & knowledge | ||
| LiveBench Data Analysis | 76.4% | 76.6% |
| LiveBench IF | 52.8% | 72.7% |
| LiveBench Language | 77.3% | 74.4% |
| LiveBench Reasoning | 77.6% | 80.0% |
| GPQA Diamond (Vals AI) | 86.4% | 88.9% |
| MMLU Pro | 86.1% | 84.3% |
| MMMU Pro | 86.0% | 83.9% |
| Maths | ||
| LiveBench Mathematics | 81.2% | 86.2% |
| ProofBench v1.1 | 21.0% | 16.0% |
| Surface Evolver Bench | 52.5% | 45.0% |
| Coding | ||
| LiveBench Agentic Coding | 56.8% | 61.4% |
| LiveBench Coding | 79.0% | 75.7% |
| Code Migration | 20.5% | 14.2% |
| CyberBench v1.1 | 67.9% | 67.8% |
| IOI | 52.5% | 39.1% |
| LiveCodeBench | 80.5% | 84.0% |
| ProgramBench | 0.0% | 0.0% |
| SWE-bench | 92.0% | 86.0% |
| Vibe Code Bench v1.1 | 30.8% | 64.8% |
| Agents | ||
| APEX-Agents | 52.8% | 47.5% |
| SkillsBench | 40.2% | 38.1% |
| Terminal-Bench-Science | 2.9% | 0.0% |
| Professional work | ||
| Finance Agent (v2) | 57.9% | 48.5% |
| EMB | 55.9% | 59.7% |
| LegalBench | 83.9% | 82.4% |
| Legal Research Bench | 45.2% | 36.1% |
| MedScribe | 88.9% | 83.9% |
| Tax Agent Bench | 31.6% | 30.9% |
| TaxEval v2 | 75.6% | 70.9% |
| Novel problems | ||
| ARC-AGI-2 | 65.8% | 42.4% |
| ARC-AGI-1 | 91.0% | 87.5% |
Tests at least two of these models have taken. Bold rows are the core tests.
