Compare / GPT-5.4 vs GPT-5.6 Sol
GPT-5.4 vs GPT-5.6 Sol
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 | GPT-5.4 | GPT-5.6 Sol |
|---|---|---|---|
| Arena (formerly LMArena)reported | Confirmed task successIPS effect estimate | -0.04 | 0.03 |
| Arena (formerly LMArena)reported | Praise over complaintIPS effect estimate | -0.01 | 0.19 |
| Arena (formerly LMArena)reported | SteerabilityIPS effect estimate | 0.02 | 0.07 |
| Arena (formerly LMArena)reported | Tool groundingIPS effect estimate | 0.00 | 0.00 |
| Arena (formerly LMArena)reported | OverallArena rating | 1,470 | 1,489 |
| Arena (formerly LMArena)reported | OverallArena rating | 1,195 | 1,256 |
| Arena (formerly LMArena)reported | OverallArena rating | 1,485 | 1,476 |
| Arena (formerly LMArena)reported | Business, management and financeArena rating | 1,482 | 1,485 |
| Arena (formerly LMArena)reported | Creative writingArena rating | 1,443 | 1,470 |
| Arena (formerly LMArena)reported | Expert promptsArena rating | 1,515 | 1,534 |
| Arena (formerly LMArena)reported | Instruction followingArena rating | 1,471 | 1,489 |
| Arena (formerly LMArena)reported | OverallArena rating | 1,475 | 1,483 |
| Arena (formerly LMArena)reported | Writing, literature and languageArena rating | 1,465 | 1,483 |
| tau2-benchreported | Consistency (pass^4)% of tasks | 21.6% | 27.8% |
| tau2-benchreported | Task success (pass^1)% of tasks | 39.4% | 46.9% |
| UGI Leaderboardreported | Requested-length error% off the requested word count | 30.0% | 1.0% |
| UGI Leaderboardreported | Style adherencescore from 0 to 1 | 0.40 | 0.37 |
| UGI Leaderboardreported | Writing scorescore out of 100 | 68.8 | 65.4 |
| Vectara Hallucination Leaderboardreported | Answer rate% of documents | 99.9% | 99.1% |
| Vectara Hallucination Leaderboardreported | Hallucination rate% of summaries | 7.0% | 12.4% |
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.