Up to four models
Gemini 3.1 Pro vs GLM-5.3-Flash
On the index (GPT-4 = 100), Gemini 3.1 Pro scores 166 and GLM-5.3-Flash 164. Side by side on the same third-party tests, with price and context. The link in your address bar shares this exact comparison.
Gemini 3.1 Pro166150–183
WeightsClosed
ReleasedFeb 2026
Price in / out$2.0 / $12
Context1.05M
Core tests taken9 of 9
GLM-5.3-Flash164148–180
WeightsOpen
ReleasedAug 2026
Price in / out$0.15 / $0.50
Context1.05M
Core tests taken6 of 9
Gemini 3.1 Pro and GLM-5.3-Flash are too close to call (3 points apart, within the uncertainty), ahead on 3 of 6 shared core tests, at 19× the price.
Test by testthird-party results
Gemini 3.1 ProGLM-5.3-FlashEvery result as a table
| Test | Gemini 3.1 Pro | GLM-5.3-Flash |
|---|---|---|
| Reasoning & knowledge | ||
| GPQA Diamond (Epoch AI) | 94.4% | 90.1% |
| GPQA Diamond (Vals AI) | 95.5% | 86.4% |
| MMLU Pro | 91.0% | 86.1% |
| MMMU Pro | 88.2% | 86.0% |
| Maths | ||
| FrontierMath (tiers 1-3) | 59.7% | 55.8% |
| AIME-style maths (OTIS mock) | 95.6% | 93.9% |
| ProofBench v1.1 | 26.0% | 21.0% |
| Chess puzzles | 51.0% | 21.0% |
| FrontierMath (tier 4) | 26.8% | 17.1% |
| Coding | ||
| Terminal-Bench 4.0 | 2.5% | 37.4% |
| Code Migration | 17.3% | 20.5% |
| IOI | 51.8% | 52.5% |
| LiveCodeBench | 88.5% | 80.5% |
| ProgramBench | 0.0% | 0.0% |
| SWE-bench | 78.8% | 92.0% |
| Vibe Code Bench 1-100 | 6.7% | 16.0% |
| Vibe Code Bench v1.1 | 32.0% | 30.8% |
| DeepSWE | 11.7% | 63.4% |
| Agents | ||
| APEX-Agents | 35.1% | 52.8% |
| SRE Bench | 0.0% | 0.8% |
| Terminal-Bench-Science | 1.4% | 2.9% |
| Mystery game puzzles | 34.0% | 8.0% |
| Professional work | ||
| Finance Agent (v2) | 43.0% | 57.9% |
| EMB | 52.6% | 55.9% |
| LegalBench | 87.4% | 83.9% |
| Legal Research Bench | 20.7% | 45.2% |
| MedScribe | 76.1% | 88.9% |
| Tax Agent Bench | 9.3% | 31.6% |
| TaxEval v2 | 72.9% | 75.6% |
| Novel problems | ||
| ARC-AGI-2 | 77.1% | 65.8% |
| ARC-AGI-1 | 98.0% | 91.0% |
Task length · METR46 sec
46 secTests at least two of these models have taken. Bold rows are the core tests.
