Compare / GPT-5.1 vs GPT-6 Astra

GPT-5.1 vs GPT-6 Astra

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

GPT-5.1
OpenAI · profile
GPT-6 Astra
OpenAI · profile
Results compared
22
Shared sources
6
BenchmarkMetricGPT-5.1GPT-6 Astra
Arena (formerly LMArena)reportedOverallArena rating 1,4161,473
Arena (formerly LMArena)reportedOverallArena rating 1,4591,465
Arena (formerly LMArena)reportedBusiness, management and financeArena rating 1,4481,483
Arena (formerly LMArena)reportedCreative writingArena rating 1,4281,453
Arena (formerly LMArena)reportedExpert promptsArena rating 1,4821,510
Arena (formerly LMArena)reportedInstruction followingArena rating 1,4501,473
Arena (formerly LMArena)reportedOverallArena rating 1,4561,478
Arena (formerly LMArena)reportedWriting, literature and languageArena rating 1,4371,469
Artificial AnalysisreportedArtificial Analysis Coding Indexindex score 49.477.1
Artificial AnalysisreportedArtificial Analysis Intelligence Indexindex score 24.752.7
Artificial AnalysisreportedGPQA Diamond% of questions 87.3%96.3%
Artificial AnalysisreportedHumanity's Last Exam% of questions 28.5%54.7%
Artificial AnalysisreportedLong-context reasoning (AA-LCR)% of questions 80.0%80.7%
Artificial AnalysisreportedTerminal-Bench 2.1% of tasks 52.4%89.9%
Artificial AnalysisreportedΤ-bench banking% of tasks 15.9%43.1%
SimpleQA Verified (Epoch AI)reportedCorrect answers% of questions 48.0%75.6%
UGI LeaderboardreportedRequested-length error% off the requested word count 113.0%0.0%
UGI LeaderboardreportedStyle adherencescore from 0 to 1 0.380.38
UGI LeaderboardreportedWriting scorescore out of 100 55.868.9
Vectara Hallucination LeaderboardreportedAnswer rate% of documents 100.0%100.0%
Vectara Hallucination LeaderboardreportedHallucination rate% of summaries 10.9%8.7%
Vending-Bench 2reportedMoney after a yearUS dollars $1,473.43$15,514.70

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.