Up to four models
GLM-5.3-Flash vs Claude Sonnet 4.6
On the index (GPT-4 = 100), GLM-5.3-Flash scores 161 and Claude Sonnet 4.6 159. Side by side on the same third-party tests, with price and context. The link in your address bar shares this exact comparison.
161145–177
WeightsOpen
ReleasedAug 2026
Price in / out$0.15 / $0.50
Context1.05M
Core tests taken6 of 9
Claude Sonnet 4.6
159143–175
WeightsClosed
ReleasedFeb 2026
Price in / out$3.0 / $15
Context1M
Core tests taken5 of 9
GLM-5.3-Flash and Claude Sonnet 4.6 are too close to call (2 points apart, within the uncertainty), ahead on 4 of 4 shared core tests, and costs 25× less.
Test by testthird-party results
Every result as a table
| Test | GLM-5.3-Flash | Claude Sonnet 4.6 |
|---|---|---|
| Reasoning & knowledge | ||
| GPQA Diamond (Epoch AI) | 90.1% | 87.4% |
| LiveBench Data Analysis | 76.4% | 78.0% |
| LiveBench IF | 52.8% | 63.2% |
| LiveBench Language | 77.3% | 76.1% |
| LiveBench Reasoning | 77.6% | 84.8% |
| GPQA Diamond (Vals AI) | 86.4% | 85.6% |
| MMLU Pro | 86.1% | 87.3% |
| MMMU Pro | 86.0% | 83.6% |
| Maths | ||
| AIME-style maths (OTIS mock) | 93.9% | 85.8% |
| LiveBench Mathematics | 81.2% | 87.0% |
| Chess puzzles | 21.0% | 23.0% |
| Coding | ||
| LiveBench Agentic Coding | 56.8% | 42.6% |
| LiveBench Coding | 79.0% | 79.3% |
| Code Migration | 20.5% | 39.9% |
| LiveCodeBench | 80.5% | 82.1% |
| ProgramBench | 0.0% | 0.5% |
| SWE-bench | 92.0% | 77.4% |
| Vibe Code Bench v1.1 | 30.8% | 51.5% |
| DeepSWE | 63.4% | 29.9% |
| Agents | ||
| APEX-Agents | 52.8% | 43.0% |
| SkillsBench | 40.2% | 49.0% |
| Mystery game puzzles | 8.0% | 16.0% |
| Professional work | ||
| Finance Agent (v2) | 57.9% | 51.0% |
| EMB | 55.9% | 60.2% |
| LegalBench | 83.9% | 82.1% |
| Legal Research Bench | 45.2% | 38.5% |
| Tax Agent Bench | 31.6% | 25.8% |
| TaxEval v2 | 75.6% | 77.1% |
| Novel problems | ||
| ARC-AGI-2 | 65.8% | 60.4% |
| ARC-AGI-1 | 91.0% | 86.5% |
Tests at least two of these models have taken. Bold rows are the core tests.