Claude Opus 4.6 vs GPT-5.4
Claude Opus 4.6 
GPT-5.4 Verdict
- Claude Opus 4.6 and GPT-5.4 are level on quality (63.1 vs 63.4).
- Claude Opus 4.6 is stronger in agents & tools, coding, human preference, knowledge, maths.
- GPT-5.4 is stronger in composite, instruction following, long context, multimodal.
- GPT-5.4 is 1.8× cheaper ($5.63 vs $10.00 per 1M blended).
- GPT-5.4 streams 3.4× faster (130 vs 38 tokens per second).
| Metric | Claude Opus 4.6 | GPT-5.4 |
|---|---|---|
| BenchLeader Index | 63.1 | 63.4 |
| Agents & tools score | 68.8 | 66.1 |
| Coding score | 62.3 | 51.6 |
| Composite score | 69.2 | 78.1 |
| Human preference score | 64.6 | 64.5 |
| Instruction following score | 53.7 | 71.7 |
| Knowledge score | 70.5 | 66.8 |
| Long context score | 65.5 | 67.6 |
| Maths score | 65.6 | – |
| Multimodal score | 63.6 | 64.0 |
| Reasoning score | 62.3 | 62.3 |
| Blended price $/M | $10.00 | $5.63 |
| Output speed | 38 tok/s | 130 tok/s |
| Time to first answer | 2.0 s | 126.4 s |
| Context window | 1M | 1.1M |
| GPQA Diamond | 90.5% | – |
| OTIS Mock AIME | 94.4% | – |
| SWE-bench Verified (Epoch) | 78.7% | – |
| Humanity's Last Exam | 19.0% | – |
| Terminal-Bench | 79.8% | 81.8% |
| SimpleBench | 67.6% | – |
| Cybench | 93.0% | – |
| Remote Labor Index | 4.2% | – |
| WeirdML | 77.9% | – |
| APEX-Agents | 32.4% | 34.9% |
| FrontierCode | 26.6% | – |
| GSO-Bench | 33.3% | – |
| Epoch Capabilities Index | 155.3 | 156.9 |
| LMArena Text | 1498 | 1466 |
| LMArena Hard Prompts | 1527 | 1488 |
| LMArena Coding | 1546 | 1514 |
| LMArena WebDev | 1537 | 1386 |
| LMArena Vision | 1311 | 1293 |
| AA Intelligence Index | 31.9 | 39.0 |
| IFBench | 53.1% | 74.0% |
| AA-LCR | 78.0% | 82.0% |
| MMMU-Pro | 75.4% | 78.4% |
| AA-Omniscience | 13.7 | 5.8 |
| Terminal-Bench Hard | 48.5% | 57.6% |
| GPQA Diamond (AA) | 89.6% | 92.0% |
| Humanity's Last Exam (AA) | 39.9% | 43.7% |
| τ²-Bench Telecom (AA) | 92.1% | 87.1% |
| HiL-Bench | 38.3% | 9.7% |
| EQ-Bench 4 | 1223 | 1272 |
| Kagi LLM Benchmark | 72.4% | 63.8% |
| SWE-bench Verified (bash only) | 75.6% | – |
| SWE-bench Verified (any scaffold) | 75.6% | – |
Data as of 2026-09-09. Best configuration of each model; every score links to its source on the model pages.