Compare / Grok 4.3 vs Grok 4.5

Grok 4.3 vs Grok 4.5

23 results from 6 sources that measured both models. Each row is in the source's own units; there is no overall winner.

Grok 4.3
xAI · profile
Grok 4.5
xAI · profile
Results compared
23
Shared sources
6
BenchmarkMetricGrok 4.3Grok 4.5
Arena (formerly LMArena)reportedOverallArena rating 1,1931,202
Arena (formerly LMArena)reportedOverallArena rating 1,4281,465
Arena (formerly LMArena)reportedBusiness, management and financeArena rating 1,4401,465
Arena (formerly LMArena)reportedCreative writingArena rating 1,4231,452
Arena (formerly LMArena)reportedExpert promptsArena rating 1,4521,488
Arena (formerly LMArena)reportedInstruction followingArena rating 1,4151,462
Arena (formerly LMArena)reportedOverallArena rating 1,4421,465
Arena (formerly LMArena)reportedWriting, literature and languageArena rating 1,4241,459
Artificial AnalysisreportedArtificial Analysis Coding Indexindex score 42.272.4
Artificial AnalysisreportedArtificial Analysis Intelligence Indexindex score 24.938.8
Artificial AnalysisreportedGPQA Diamond% of questions 90.1%93.1%
Artificial AnalysisreportedHumanity's Last Exam% of questions 37.2%42.7%
Artificial AnalysisreportedLong-context reasoning (AA-LCR)% of questions 75.0%79.3%
Artificial AnalysisreportedSciCode% of problems 48.3%55.0%
Artificial AnalysisreportedTerminal-Bench 2.1% of tasks 39.7%81.6%
Artificial AnalysisreportedTerminal-Bench 4.0% of tasks 0.0%10.6%
Artificial AnalysisreportedΤ-bench banking% of tasks 12.4%42.1%
GDP.pdfreportedProfessional document tasks% of rubric 8.0%14.0%
SimpleQA Verified (Epoch AI)reportedCorrect answers% of questions 33.2%48.3%
UGI LeaderboardreportedRequested-length error% off the requested word count 21.0%21.0%
UGI LeaderboardreportedStyle adherencescore from 0 to 1 0.340.32
UGI LeaderboardreportedWriting scorescore out of 100 57.760.0
Vending-Bench 2reportedMoney after a yearUS dollars $35.26$3,887.43

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.