Compare / Claude Sonnet 4.6 vs GPT-5.4
Claude Sonnet 4.6 vs GPT-5.4
20 results from 4 sources that measured both models. Each row is in the source's own units; there is no overall winner.
| Benchmark | Metric | Claude Sonnet 4.6 | GPT-5.4 |
|---|---|---|---|
| Arena (formerly LMArena)reported | OverallArena rating | 1,478 | 1,470 |
| Arena (formerly LMArena)reported | OverallArena rating | 1,201 | 1,195 |
| Arena (formerly LMArena)reported | OverallArena rating | 1,461 | 1,485 |
| Arena (formerly LMArena)reported | Business, management and financeArena rating | 1,482 | 1,482 |
| Arena (formerly LMArena)reported | Creative writingArena rating | 1,449 | 1,443 |
| Arena (formerly LMArena)reported | Expert promptsArena rating | 1,506 | 1,515 |
| Arena (formerly LMArena)reported | Instruction followingArena rating | 1,475 | 1,471 |
| Arena (formerly LMArena)reported | OverallArena rating | 1,472 | 1,475 |
| Arena (formerly LMArena)reported | Writing, literature and languageArena rating | 1,455 | 1,465 |
| OpenHands Indexreported | Average% resolved | 44.5% | 64.3% |
| OpenHands Indexreported | Front end (SWE-Bench Multimodal)% resolved | 30.9% | 36.8% |
| OpenHands Indexreported | Greenfield (Commit0)% resolved | 50.0% | 56.2% |
| OpenHands Indexreported | Information gathering (GAIA)% resolved | 13.3% | 82.4% |
| OpenHands Indexreported | Issue resolution (SWE-Bench)% resolved | 74.4% | 75.6% |
| OpenHands Indexreported | Testing (SWT-Bench)% resolved | 54.0% | 70.4% |
| UGI Leaderboardreported | Requested-length error% off the requested word count | 2.0% | 30.0% |
| UGI Leaderboardreported | Style adherencescore from 0 to 1 | 0.41 | 0.40 |
| UGI Leaderboardreported | Writing scorescore out of 100 | 64.5 | 68.8 |
| Vectara Hallucination Leaderboardreported | Answer rate% of documents | 99.9% | 99.9% |
| Vectara Hallucination Leaderboardreported | Hallucination rate% of summaries | 10.6% | 7.0% |
Bold green marks the better value on that metric. Where a source reports ranges that overlap, the difference may not be meaningful; see the benchmark page for ranges.