GPT-5.4 Pro vs Grok 4.5
GPT-5.4 Pro 
Grok 4.5 Verdict
- GPT-5.4 Pro and Grok 4.5 are level on quality (64.2 vs 63.4).
- GPT-5.4 Pro is stronger in instruction following, multimodal, reasoning.
- Grok 4.5 is stronger in knowledge, agents & tools, coding, composite, human preference, long context.
- Grok 4.5 is 23× cheaper ($3.00 vs $67.50 per 1M blended).
- Grok 4.5 streams 55.7× faster (56 vs 1 tokens per second).
| Metric | GPT-5.4 Pro | Grok 4.5 |
|---|---|---|
| BenchLeader Index | 64.2 | 63.4 |
| Instruction following score | 67.3 | – |
| Knowledge score | 73.2 | 76.0 |
| Multimodal score | 69.8 | 64.8 |
| Reasoning score | 75.1 | 68.9 |
| Agents & tools score | – | 58.6 |
| Coding score | – | 59.9 |
| Composite score | – | 66.0 |
| Human preference score | – | 66.9 |
| Long context score | – | 66.2 |
| Blended price $/M | $67.50 | $3.00 |
| Output speed | 1 tok/s | 56 tok/s |
| Time to first answer | 6.7 s | 12.7 s |
| Context window | 1.1M | 500k |
| Humanity's Last Exam | 44.3% | – |
| SimpleBench | 74.1% | 70.0% |
| WeirdML | – | 46.4% |
| APEX-Agents | – | 34.2% |
| FrontierCode | – | 42.4% |
| Epoch Capabilities Index | 158.9 | 153.9 |
| LMArena Text | – | 1471 |
| LMArena Hard Prompts | – | 1495 |
| LMArena Coding | – | 1523 |
| LMArena WebDev | – | 1556 |
| LMArena Vision | – | 1291 |
| LMArena Agent | – | 3.9 |
| LiveBench | – | 75.8% |
| LiveBench Reasoning | – | 87.2% |
| LiveBench Coding | – | 68.6% |
| LiveBench Agentic Coding | – | 56.5% |
| LiveBench Mathematics | – | 90.8% |
| LiveBench Data Analysis | – | 73.0% |
| LiveBench Language | – | 82.8% |
| AA Intelligence Index | – | 39.1 |
| AA-LCR | – | 79.3% |
| MMMU-Pro | – | 80.4% |
| AA-Omniscience | – | 25.3 |
| GPQA Diamond (AA) | – | 93.1% |
| Humanity's Last Exam (AA) | – | 42.7% |
| SciCode (AA) | – | 55.0% |
| MultiChallenge | 69.2% | – |
| VISTA | 53.9% | – |
| MultiNRC | 62.3% | – |
| TutorBench | 56.6% | – |
| Kagi LLM Benchmark | – | 83.5% |
Data as of 2026-09-09. Best configuration of each model; every score links to its source on the model pages.