Benchmarks people trust to tell if an LLM is frontier

I asked 56 people in AI which benchmarks they'd trust most to indicate if an LLM is frontier:
VotesBenchmarkCreatorsArea
25Terminal-Bench 4Alex Shaw (@alexgshaw)Andy Konwinski (@andykonwinski)Mike A. Merrill (@Mike_A_Merrill)Alex ShawS., Andy KonwinskiK., Mike A. MerrillM.Coding
20Terminal-Bench ScienceSteven Dillmann (@StevenDillmann)Ludwig Schmidt (@lschmidt3)Sanmi Koyejo (@sanmikoyejo)Steven DillmannD., Ludwig SchmidtS., Sanmi KoyejoK.Science
13FrontierCodeEric Lu (@penlume)Ben Pan (@ybenpan)Deniz Birlikci (@denizbirlikci)Silas Alberti (@silasalberti)Eric LuL., Ben PanP., Deniz BirlikciB., Silas AlbertiA.Coding
10FrontierSWEEvan Chu (@evan_j_chu)Rajan Agarwal (@_rajanagarwal)Justus Mattern (@MatternJustus)Rishyanth Kondra (@ardnoKhtnayhsiR)Sanket Mhatre (@sanket__mhatre)Akshit Kumar (@akshitkr)Evan ChuC., Rajan AgarwalA., Justus MatternM., Rishyanth KondraK., Sanket MhatreM., Akshit KumarK.Coding
8Humanity's Last ExamLong Phan (@longphann_)Dan Hendrycks (@hendrycks)Alice GattiZiwen Han (@ziwen_h)Nathaniel Li (@natliml)Long PhanP., Dan HendrycksH., Alice GattiG., Ziwen HanH., Nathaniel LiL.Expert knowledge
7Vending-BenchAxel Backlund (@axelbacklund)Lukas Petersson (@lukaspet)Axel BacklundB., Lukas PeterssonP.Business
5AutomationBenchDaniel Shepard (@danielwshepard)Robin Salimans (@SalimansRobin)Anna Marie Clifton (@TweetAnnaMarie)Daniel ShepardS., Robin SalimansS., Anna Marie CliftonC.Business
5DeepSWEWenqi (Kevin) Huang (@winkey_h)Charley Lee (@charleyslee)Leonard Tng (@LeonardMainnet)Serena Ge (@serenaa_ge)Wenqi (Kevin) HuangH., Charley LeeL., Leonard TngT., Serena GeG.Coding
5PostTrainBenchBen Rank (@full__rank)Hardik Bhatnagar (@hrdkbhatnagar)Ameya Prabhu (@AmyPrb)Shira Eisenberg (@shiraeis)Karina Nguyen (@karinanguyen)Matthias Bethge (@MatthiasBethge)Maksym Andriushchenko (@maksym_andr)Ben RankR., Hardik BhatnagarB., Ameya PrabhuP., Shira EisenbergE., Karina NguyenN., Matthias BethgeB., Maksym AndriushchenkoA.ML engineering
4AA-BriefcaseMicah Hill-Smith (@_micah_h)George Cameron (@grmcameron)Micah Hill-SmithH., George CameronC.Knowledge work
4Agents' Last ExamYiyou Sun (@YiyouSun)Dawn Song (@dawnsongtweets)Xinyang Han (@David_Han66)Weichen Zhang (@WeichenZhang339)Yiyou SunS., Dawn SongS., Xinyang HanH., Weichen ZhangZ.Knowledge work
4FrontierCS 2Qiuyang Mang (@MangQiuyang)Wenhao Chai (@wenhaocha1)Runyuan He (@RunyuanHe)Dongruixuan Li (@dongruixuan)Zihan Zheng (@ZihanZheng71803)Yujun He (@Yujun__He)Alvin Cheung (@alvinkcheung)Joseph E. Gonzalez (@profjoeyg)Qiuyang MangM., Wenhao ChaiC., Runyuan HeH., Dongruixuan LiL., Zihan ZhengZ., Yujun HeH., Alvin CheungC., Joseph E. GonzalezG.Coding
4OSWorld 2Mengqi Yuan (@yuan_mengq43669)Zilong Zhou (@adlsdztony1)Tao Yu (@taoyds)Tianbao Xie (@TianbaoX)Mengqi YuanY., Zilong ZhouZ., Xinzhuang XiongX., Tao YuY., Tianbao XieX.Computer use
3ARC-AGI-3Greg Kamradt (@GregKamradt)François Chollet (@fchollet)Mike Knoop (@mikeknoop)Hunter HenryGreg KamradtK., François CholletC., Mike KnoopK., Hunter HenryH.Reasoning
3FrontierMathElliot Glazer (@ElliotGlazer)Tamay Besiroglu (@tamaybes)Ege Erdil (@EgeErdil2)Elliot GlazerG., Tamay BesirogluB., Ege ErdilE.Math
3Harvey LABNiko Grupen (@nikogrupen)Gabe Pereyra (@gabepereyra)Julio Pereyra (@ItsJulioPereyra)Spencer Poff (@spencerpoff)Niko GrupenG., Gabe PereyraP., Julio PereyraP., Spencer PoffP.Legal
3ProgramBenchJohn Yang (@jyangballin)Kilian Lieret (@KLieret)Jeffrey Jian Ma (@18jeffreyma)Diyi Yang (@Diyi_Yang)Ofir Press (@OfirPress)John YangY., Kilian LieretL., Jeffrey Jian MaM., Diyi YangY., Ofir PressP.Coding
3SWE-sweepKilian Lieret (@KLieret)Jeffrey Jian Ma (@18jeffreyma)Ofir Press (@OfirPress)John Yang (@jyangballin)Kilian LieretL., Jeffrey Jian MaM., Ofir PressP., John YangY.Coding
2GDPxMax Kan (@maxkan)Andrew Megalaa (@AndrewMegalaa)Dylan Patel (@dylan522p)Abhijay Rana (@abhijaymrana)Max KanK., Andrew MegalaaM., Dylan PatelP., Abhijay RanaR.Multi-domain
2BrowseCompJason Wei (@_jasonwei)Zhiqing Sun (@EdwardSun0909)Hyung Won Chung (@hwchung27)Mia (Amelia) Glaese (@mia_glaese)Jason WeiW., Zhiqing SunS., Hyung Won ChungC., Mia (Amelia) GlaeseG.Web research
2ChartographyChris MuttyEdwin Chen (@echen)SurgeSuhaas GarreG., Chris MuttyM., Sushant MehtaM., Edwin ChenC.Visual reasoning
2CursorBenchNaman Jain (@StringChaos)Naman JainJ.Coding
2ExploitBenchSeunghyun Lee (Xion) (@0x10n)David Brumley (@thedavidbrumley)Seunghyun Lee (Xion)(., David BrumleyB.Cyber
2GDP.pdfEdwin Chen (@echen)SurgeSuhaas GarreG., Emily RitchieR., Sushant MehtaM., Edwin ChenC.Knowledge work
2GDPval-AATejal Patwardhan (@tejalpatwardhan)Micah Hill-Smith (@_micah_h)George Cameron (@grmcameron)Declan Jackson (@not_benchmaxing)Tejal PatwardhanP., Micah Hill-SmithH., George CameronC., Declan JacksonJ.Knowledge work

Counts are net: a few benchmarks got negative votes.

GDPx came out the day before I sent the survey.

GDPx was built with 14 data partners: Arcverex, Aria, Arithmetic, Beautiful Machines, Chakra, Datacurve, Emulated, Ergodex, Extrapolate, Fleet, Latch, Living Context, Third Moment, and ValenceForge.

The most common comment was that people don't really trust benchmarks at all.

Also mentioned (one vote each)

Also picked but outside the rules (indexes, human-voted boards, and bench suites)

Voted by 56 people
Originally posted on X on Nov 15, 2026.