GPT-5.4 Pro vs Kimi K3
GPT-5.4 Pro 
Kimi K3 Verdict
- Kimi K3 leads on quality: 66.5 vs 64.2.
- GPT-5.4 Pro is stronger in instruction following, knowledge, multimodal, reasoning.
- Kimi K3 is stronger in agents & tools, coding, composite, human preference, long context, maths.
- Kimi K3 is 11× cheaper ($6.00 vs $67.50 per 1M blended).
- Kimi K3 streams 36.9× faster (37 vs 1 tokens per second).
| Metric | GPT-5.4 Pro | Kimi K3 |
|---|---|---|
| BenchLeader Index | 64.2 | 66.5 |
| Instruction following score | 67.3 | – |
| Knowledge score | 73.2 | 68.1 |
| Multimodal score | 69.8 | 65.1 |
| Reasoning score | 75.1 | – |
| Agents & tools score | – | 67.5 |
| Coding score | – | 64.9 |
| Composite score | – | 74.1 |
| Human preference score | – | 71.2 |
| Long context score | – | 71.1 |
| Maths score | – | 78.1 |
| Blended price $/M | $67.50 | $6.00 |
| Output speed | 1 tok/s | 37 tok/s |
| Time to first answer | 6.7 s | 57.7 s |
| Context window | 1.1M | 1.0M |
| Humanity's Last Exam | 44.3% | – |
| SimpleBench | 74.1% | – |
| SciCode | – | 58.7% |
| APEX-Agents | – | 39.3% |
| FrontierCode | – | 44.2% |
| ProofBench | – | 87.0% |
| Epoch Capabilities Index | 158.9 | 157.6 |
| LiveBench | – | 79.2% |
| LiveBench Reasoning | – | 90.7% |
| LiveBench Coding | – | 81.5% |
| LiveBench Agentic Coding | – | 62.2% |
| LiveBench Mathematics | – | 84.4% |
| LiveBench Data Analysis | – | 78.7% |
| LiveBench Language | – | 85.5% |
| AA Intelligence Index | – | 43.8 |
| AA-LCR | – | 88.7% |
| MMMU-Pro | – | 80.5% |
| AA-Omniscience | – | 19.7 |
| GPQA Diamond (AA) | – | 93.5% |
| Humanity's Last Exam (AA) | – | 46.9% |
| SciCode (AA) | – | 59.5% |
| LegalBench | – | 86.0% |
| CorpFin | – | 71.6% |
| TaxEval | – | 75.7% |
| Terminal-Bench 2.1 (Vals) | – | 80.9% |
| SWE-bench (Vals) | – | 93.4% |
| GPQA Diamond (Vals) | – | 92.9% |
| Vals Index | – | 57.8 |
| MultiChallenge | 69.2% | – |
| VISTA | 53.9% | – |
| MultiNRC | 62.3% | – |
| TutorBench | 56.6% | – |
| EQ-Bench 4 | – | 1339 |
Data as of 2026-09-09. Best configuration of each model; every score links to its source on the model pages.