GPT-5.4 vs Grok 4.5
GPT-5.4 
Grok 4.5 Verdict
- GPT-5.4 and Grok 4.5 are level on quality (63.4 vs 63.4).
- GPT-5.4 is stronger in agents & tools, composite, instruction following, long context.
- Grok 4.5 is stronger in coding, human preference, knowledge, multimodal, reasoning.
- Grok 4.5 is 1.9× cheaper ($3.00 vs $5.63 per 1M blended).
- GPT-5.4 streams 2.3× faster (130 vs 56 tokens per second).
| Metric | GPT-5.4 | Grok 4.5 |
|---|---|---|
| BenchLeader Index | 63.4 | 63.4 |
| Agents & tools score | 66.1 | 58.6 |
| Coding score | 51.6 | 59.9 |
| Composite score | 78.1 | 66.0 |
| Human preference score | 64.5 | 66.9 |
| Instruction following score | 71.7 | – |
| Knowledge score | 66.8 | 76.0 |
| Long context score | 67.6 | 66.2 |
| Multimodal score | 64.0 | 64.8 |
| Reasoning score | 62.3 | 68.9 |
| Blended price $/M | $5.63 | $3.00 |
| Output speed | 130 tok/s | 56 tok/s |
| Time to first answer | 126.4 s | 12.7 s |
| Context window | 1.1M | 500k |
| Terminal-Bench | 81.8% | – |
| SimpleBench | – | 70.0% |
| WeirdML | – | 46.4% |
| APEX-Agents | 34.9% | 34.2% |
| FrontierCode | – | 42.4% |
| Epoch Capabilities Index | 156.9 | 153.9 |
| LMArena Text | 1466 | 1471 |
| LMArena Hard Prompts | 1488 | 1495 |
| LMArena Coding | 1514 | 1523 |
| LMArena WebDev | 1386 | 1556 |
| LMArena Vision | 1293 | 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 | 39.0 | 39.1 |
| IFBench | 74.0% | – |
| AA-LCR | 82.0% | 79.3% |
| MMMU-Pro | 78.4% | 80.4% |
| AA-Omniscience | 5.8 | 25.3 |
| Terminal-Bench Hard | 57.6% | – |
| GPQA Diamond (AA) | 92.0% | 93.1% |
| Humanity's Last Exam (AA) | 43.7% | 42.7% |
| SciCode (AA) | – | 55.0% |
| τ²-Bench Telecom (AA) | 87.1% | – |
| HiL-Bench | 9.7% | – |
| EQ-Bench 4 | 1272 | – |
| Kagi LLM Benchmark | 63.8% | 83.5% |
Data as of 2026-09-09. Best configuration of each model; every score links to its source on the model pages.