
[2026幎8æç] ç¡æã®AIã©ã³ãã³ã°ã¯ã幎å150ååã®ããžãã¹ã«ãªã£ãŠãããArenaã®çŸåšå°
å幎åã«æžããèšäºãã
ããŸã ã«æ¯æãã¡ã°ãèªãŸããŠããŸãã
AIã¢ãã«ã®æ Œä»ãæ©é¢ãArenaã玹ä»ããèšäºã§ãã
ä»çµã¿ã説æããŠããªãŒããŒããŒãã®èªã¿æ¹ãæžããŠã
ããã¯ããŒã¯ããŠããããšç· ããã
ãã®èšäºãæžãããšãã®ç§ã¯ã
ã²ãšã€å€§ããªããšãèŠèœãšããŠããŸããã
ãã®ã©ã³ãã³ã°ãäœã£ãŠããçµç¹ãã
äœã§é£ã¹ãŠããã®ããæžããŠããªãã£ãã
2026幎6æãTechCrunchããããå ±ããŸããã
Arenaã®å¹Žæç®åçã¯1åãã«ã
æ¥æ¬åã§ããã150ååã§ãã
ç¡æã®ã©ã³ãã³ã°ã ãšæã£ãŠãããã®ã¯ã
ç¡æã®é客è£
眮ã§ããã
åç»æŽŸã®ããã¯ãã¡ããž
ãã®èšäºã®äœçœ®ã¥ã
ããã ââ 2026幎8ææç¹ã®Arenaã®çŸåšå°ããéå¶æ§é ãŸã§å«ããŠæŽæ°ãã
察象 ââ ãªãŒããŒããŒãã®æ°åãæ ¹æ ã«ã瀟å ã§ã¢ãã«éžå®ã®èª¬æãããªããã°ãªããªã人ãäŒç»ãPdMãããŒã±ãæ ã·ã¹
ãŽãŒã« ââ ã©ã³ãã³ã°ã®æ°åããã©ããŸã§ä¿¡ããŠã©ãããçããã®ç·ãåŒããããã«ãªã
2026幎2æã«æžããå ¥éç·šã¯ãã¡ãã§ããä»çµã¿ãã®ãã®ãç¥ãããæ¹ã¯ãå ã«ãã¡ããã©ããã
æéããªã人åããŸãšã
Arenaã®å¹Žæç®åçã¯1åãã«ãåçšãµãŒãã¹ã®éå§ããããã8ã¶æã§ã®å°éã§ã
åçæºã¯ç¡æã®ãªãŒããŒããŒãã§ã¯ãªããã¢ãã«ã©ããšäŒæ¥ã«å£²ã£ãŠããæåã®è©äŸ¡åæã§ã
è©äŸ¡ã«åå ãããŠãŒã¶ãŒã®å€ãã¯ãæªçºè¡šã¢ãã«ã«å ã«è§ŠããããŠæ¥ãŠããŸããã®æ§é ããç¡æã®æç¥šãæç«ãããŠããŸã
2025幎ã®è«æã§ãäžéšã®ãããã€ããéå ¬éãã¹ããç¹°ãè¿ããŠæè¯ã¹ã³ã¢ã ããæåºã§ããä»çµã¿ãææãããŠããŸããNeurIPS 2025æ¡æ
ãã ãããããäžæ£ã ãšåŒã¶ã®ã¯éããŸããæ§é ã®è©±ã§ãã£ãŠãæªæã®è©±ã§ã¯ãããŸãã
å®åã§ã¯ãé äœã§ã¯ãªãæç¥šæ°ãšæŽæ°æ¥ãèŠãŠãã ãããçç±ã¯æ¬æã§æžããŸã
å€ãã£ãããš1ãArenaã¯äŒç€Ÿã«ãªãã幎å150ååã«ãªã£ã
ãŸãæ°åãã䞊ã¹ãŸãããã¹ãŠTechCrunchã®å ±éãšãå ¬éãããŠããè³éèª¿éæ å ±ã§ãã
2023幎ãã«ãªãã©ã«ãã¢å€§åŠããŒã¯ã¬ãŒæ ¡ã®ç ç©¶ãããžã§ã¯ããšããŠèªç
2025幎4æãArena Intelligence Inc. ãšããŠæ³äººå
2025幎5æãè©äŸ¡é¡6åãã«ã§1åãã«ã®ã·ãŒãã©ãŠã³ã
2025幎9æãæåãµãŒãã¹ã®AI Evaluationsãéå§
2026幎1æãè©äŸ¡é¡17åãã«ã§1å5,000äžãã«ã®ã·ãªãŒãºAããã®æç¹ã®å¹Žæç®åçã¯3,000äžãã«
2026幎1æ28æ¥ããã©ã³ãåãArenaã«å€æŽãããã¡ã€ã³ãarena.aiãžç§»è¡
2026幎6æã幎æç®åç1åãã«ã«å°éã环èšèª¿éé¡ã¯2å5,000äžãã«
5ã¶æã§3,000äžãã«ãã1åãã«ã§ã ã²ããããïŒïŒ
å ±å嵿¥è ã¯Anastasios Angelopoulosæ°ïŒCEOïŒãWei-Lin Chiangæ°ïŒCTOïŒããããŠDatabrickså ±å嵿¥è ã§ããŒã¯ã¬ãŒææã®Ion Stoicaæ°ã
åºå žã¯ãã¡ãã§ãã
å€ãã£ãããš2ã誰ãéãæã£ãŠããã®ã
ããããã®èšäºã®æ¬é¡ã§ãã
ãªãŒããŒããŒãã¯ç¡æã§ããæç¥šãããŠãŒã¶ãŒãç¡æã§ããããã1åãã«ã¯èª°ãæã£ãŠããã®ãã
ã¢ãã«ã©ããšäŒæ¥ã§ã
Arenaã2025幎9æã«å§ããAI Evaluationsã¯ãã³ãã¥ããã£ããéããè©äŸ¡ããŒã¿ãããšã«ããæ·±æãã®æ§èœåæããã¢ãã«éçºäŒæ¥ãšäžè¬äŒæ¥ã«æäŸãããµãŒãã¹ã§ãã
ã€ãŸããããããæ§é ã«ãªã£ãŠããŸãã
ç¡æã®ãªãŒããŒããŒããã倧éã®è©äŸ¡è ãéãã
è©äŸ¡è ãæç¥šããããšã§ãã¢ãã«ã®æ¯èŒããŒã¿ãèç©ãã
ãã®ããŒã¿ãåæãããã®ããã¢ãã«ãäœã£ãŠããåŽã«å£²ã
CEOã®Angelopoulosæ°ã¯TechCrunchã«ããèªã£ãŠããŸãã
people still see us as an open source project
ïŒããŸã ã«å€ãã®äººããç§ãã¡ããªãŒãã³ãœãŒã¹ãããžã§ã¯ãã ãšæã£ãŠããïŒ
åºå žïŒArena, the AI leaderboard everyone uses, is now a $100M business, TechCrunch, 29 June 2026
ç§ãããæã£ãŠããåŽã§ããå幎åã®èšäºã«ããæžããŠããŸããã
ãã®æ§é ãã©ããã§èŠãããšããããŸããã
ç¡æã§äŸ¡å€ãé ã£ãŠäººãéããéãŸã£ã人ããçãŸããããŒã¿ãå¥ã®çžæã«å£²ãã
æ€çŽ¢ãšã³ãžã³ããSNSããäŸ¡æ Œæ¯èŒãµã€ãããå šéšãã®åœ¢ã§ãã
æªããšèšã£ãŠããã®ã§ã¯ãããŸããããããéåžžã«ããèšèšãããäºæ¥ã ãšæããŸããç¡æã®æç¥šãç¶ããããããŒã¿ãå¢ããããŒã¿ãå¢ããã»ã©åæã®äŸ¡å€ãäžãããåå ¥ãé£ããã
ãã ãããã®ã©ã³ãã³ã°ãèŠããšãã®åæãå€ãããŸã
å€ãã£ãããš3ãARRãšããèšèã®äœãæžã
现ãã話ã«èŠããŠãå®åã§ã¯ãã£ãã倧äºãªç¹ããããŸãã
Arenaã¯1åãã«ãšããæ°åãARRãšåŒãã§ããŸãããAngelopoulosæ°ã¯TechCrunchã«å¯ŸããŠãå瀟ã®èª²éã¯æ¶è²»éã«ããšã¥ããã®ã§ãç¶ç¶çãªåçã§ã¯ãªããšæç¢ºã«ããŠããŸãã
AIã®äžçã§ã¯ããã®æã®æ°åãããåºãŠããŸãã幎æç®åçãšæžãããŠããŠãããããå®å®ããŠç¹°ãè¿ãããå¥çŽãªã®ãã䜿ã£ãåã ãæãåŸé課éã®ç¬é颚éãªã®ãã§ãæå³ããŸã£ããéãã
ããã¯ãAIäŒæ¥ã®æ°åãèªããšãã®äžè¬åãšããŠã䜿ããŸã
èªç€Ÿãå°å ¥ãæ€èšããŠããAIãã³ããŒããæ¥æé·äžã®ARRãèŠããŠãããšãããããåŸé課éã®åèšãªã®ããèãã ãã§ãçžæã®çãæ¹ãå€ãããŸãã
å€ãã£ãããš4ããªãç¡æã§æç¥šãã人ãããã®ã
ããã¯ãç§ãå幎åã«ã¡ãããšæžããŠããªãã£ãéšåã§ãã
Arenaã®ãŠãŒã¶ãŒã¯ãåæã ãã§æç¥šããŠããããã§ã¯ãããŸããã
TechCrunchã®èšäºã«ã¯ãè©äŸ¡è ããã©ãããã©ãŒã ã«éãŸãçç±ãšããŠãææ°ã®ããã°ãã°æªçºè¡šã®ã¢ãã«ã«æ©ãè§ŠããããããšãæããããŠããŸãã
ã€ãŸãããããã亀æãæç«ããŠããŸãã
ãŠãŒã¶ãŒã¯ããŸã äžã«åºãŠããªãã¢ãã«ãè§Šãã
Arenaã¯ããã®å¯ŸäŸ¡ãšããŠæ¯èŒããŒã¿ãåŸã
ã¢ãã«ã©ãã¯ãå ¬éåã«å€§èŠæš¡ãªäººéè©äŸ¡ãåŸããã
äžè ãšãåŸããããããã§ããèšèšã§ãã
ãããŠããã®èšèšããæ¬¡ã«èª¬æããåé¡ã®å ¥ãå£ã§ããããŸã
å幎åããææãããŠãããæ§é ã®åé¡
2025幎ã«çºè¡šãããNeurIPS 2025ã®Datasets and Benchmarks Trackã«æ¡æãããè«æããããŸããThe Leaderboard IllusionãCohere Labsãã¹ã¿ã³ãã©ãŒããããªã³ã¹ãã³ãMITããŠã©ãŒã¿ãŒã«ãŒå€§ãªã©ã®å ±åç ç©¶ã§ãã
243ã¢ãã«ã42ãããã€ãã200äžä»¶ã®å¯ŸæŠãåæããŠããŸãã
ææãããŠããã®ã¯ãäž»ã«4ç¹ã§ãã
1. éå ¬éãã¹ããšãã¹ã³ã¢ã®åãäžã
äžéšã®ãããã€ãã¯ãå ¬éåã«è€æ°ã®ããªã¢ã³ããéå ¬éã§ãã¹ãã§ããçµæãæªããã°ã¹ã³ã¢ãåãäžããããã
è«æã¯ããæžããŠããŸãã
we identify 27 private LLM variants tested by Meta
ïŒMetaã27ã®éå ¬éLLMããªã¢ã³ãããã¹ãããŠããããšãç¹å®ããïŒ
åºå žïŒThe Leaderboard Illusion, Singh et al., arXiv:2504.20879
Llama-4ã®ãªãªãŒã¹åã®æéã§ããvisionåãã®ãªãŒããŒããŒãã®åãå«ãããšãããã«16ä»¶ãå ãããšãããŠããŸãã
äžæ¹ãå°èŠæš¡ãªã¹ã¿ãŒãã¢ããã¯1ã€ã2ã€ã芳枬æéäžãåŠè¡ã©ãã«ããéå ¬éãã¹ãã¯1ä»¶ã確èªãããªãã£ããšè«æã¯æžããŠããŸãã
2. ããŒã¿ã¢ã¯ã»ã¹ã®åã
察æŠã®å²ãåœãŠã«ãå·®ããããŸããè«æã®æšèšã§ã¯ãGoogleãšOpenAIãããããå šããŒã¿ã®19.2ããŒã»ã³ããš20.4ããŒã»ã³ããåãåã£ãŠããã察ããŠã83ã®ãªãŒãã³ãŠã§ã€ãã¢ãã«ã®åèšã29.7ããŒã»ã³ãã§ãã
3. ãã®ããŒã¿ãæ§èœã«å¹ã
è«æã¯ãArenaã®ããŒã¿ãžã®ã¢ã¯ã»ã¹ãå®éã«æ§èœåäžãããããããšãæ€èšŒããŠããŸããéããã远å ããŒã¿ã§ããArenaäžã®ååžã«ãããçžå¯Ÿæ§èœãæå€§112ããŒã»ã³ãåäžããããšããæšèšã§ãã
ã€ãŸããããŒã¿ãå€ãããããã»ã©ãArenaã§åŒ·ããªããããŠåŒ·ãã»ã©ããŸãå€ã察æŠãå²ãåœãŠãããã
4. ãµã€ã¬ã³ããªåé€
ã¢ãã«ããªãŒããŒããŒãããéãã«åãé€ãããããšãããããã®éçšã®éææ§ãäœãããšãææãããŠããŸãã
ãã ãããããäžæ£ãšåŒã¶ã®ã¯éããŸã
ããã¯äžå¯§ã«æžããŸãã
è«æãææããŠããã®ã¯ãç¹å®ã®ããªã·ãŒããçµ±èšçãªåæãå£ããŠãããšããæ§é ã®åé¡ã§ãã誰ããåãã€ãããšãã話ã§ã¯ãããŸããã
Arenaã®ã©ã³ãã³ã°ã¯ãBradley-Terryã¢ãã«ãšããæ çµã¿ã§èšç®ãããŠããŸãããã®æ çµã¿ã¯ãæåºãããã¢ãã«ãæ£æçã«éžã°ããŠããªãããšãåæã«ããŠããŸãã
ãšãããã10åãã¹ãããŠæè¯ã®1ã€ã ããåºãããããã§ãããšãåæã厩ããŸãã
ãããŠããã®ããæ¹ãå¯èœã ãšç¥ã£ãŠãããããã€ããšãç¥ããªãã£ããããã€ããããè«æãåé¡ã«ããŠããã®ã¯ãããã§ãã
ArenaåŽã¯ææã«åè«ããŠãããè°è«ã¯ç¶ç¶ããŠããŸããç§ã¯ã©ã¡ããæ£ããããå€å®ããç«å Žã«ãããŸããã
æžããã®ã¯ããã®æ§é ãç¥ã£ãããã§æ°åãèŠãŠãã ãããšããããšã ãã§ãã
ããããã©ãèªããã5ã€ã®ã«ãŒã«
å®åã§äœ¿ãã圢ã«ããŸãã
ã«ãŒã«1ãé äœã§ã¯ãªããã¹ã³ã¢å·®ãèŠã
äžäœã¢ãã«ã®ã¬ãŒãã£ã³ã°ã¯ããã°ãã°æ°ãã€ã³ãããåæ°ãã€ã³ãã®å·®ã§äžŠãã§ããŸãããã®å·®ã¯ãä¿¡é Œåºéã®äžã«åãŸã£ãŠããããšããããŸãã
1äœãš3äœãå®è³ªçã«åãããšããç¶æ³ã¯æ®éã«èµ·ããŸã
é äœã ããåŒçšãããšããããèœãšããŸãã
ã«ãŒã«2ãæç¥šæ°ãå¿ ãèŠã
æ°ããåºãã¢ãã«ã¯ãå¯ŸæŠæ°ãå°ãªãç¶æ ã§é«ãã¹ã³ã¢ãåºãããšããããŸãã
ã¹ã³ã¢ã®æšªã«ããæç¥šæ°ãèŠãŠãæ¡ãä»ãšæã£ãŠãããã確èªããŠãã ãããæ¡ã1ã€å°ãããªãããã®æ°åã¯ãŸã æ«å®ã§ãã
ã«ãŒã«3ãç·åã§ã¯ãªããçšéå¥ã®ããŒããèŠã
Arenaã¯ããã¹ããã³ãŒãã£ã³ã°ãããžã§ã³ãç»åçæããããŠé·æéã®ã¯ãŒã¯ãããŒãæ±ãAgent ModeãŸã§ãè€æ°ã®ããŒããæã£ãŠããŸãã
ç·åã®é äœã¯ãå¹³åçãªå¥œãŸããã衚ããŠããŸããããªãã®æ¥åãå¹³åçã§ããããšã¯ããŸããããŸããã
ã«ãŒã«4ã人éã®å¥œã¿ã枬ã£ãŠããããšããåæãå¿ããªã
Arenaãæž¬ã£ãŠããã®ã¯ã人éãèªãã§è¯ããšæãããã©ããã§ãã
æ£ç¢ºãã§ã¯ãããŸãããäºå®ãšããŠæ£ãããã©ããã¯ãæç¥šè ãæ€èšŒããŠããªãå Žåãå€ãã
èªã¿ãããåã¯ãArenaã§ã¯åã¡ãŸã
ã ãããäºå®ã®æ£ç¢ºããèŠãæ¥åã§ã¯ãArenaã ãã§æ±ºããŠã¯ãããŸããã
ã«ãŒã«5ãçšè°æžã«ã¯ãæ¥ä»ãæžã
ãããäžçªå®åçã§ãã
ãªãŒããŒããŒãã¯é »ç¹ã«æŽæ°ãããŸããå幎åã®é äœãåŒçšããŠæææ±ºå®ãããšãããšã§èª¬æã§ããªããªããŸãã
瀟å è³æã«æžããšãã¯ãå¿ ããã®åœ¢ã«ããŠãã ããã
åºå žïŒArenaïŒarena.aiïŒ
åç §ããããŒãå
åç §æ¥
ã¹ã³ã¢ãšæç¥šæ°
æ¥ä»ãšæç¥šæ°ãæžããŠããè³æã¯ãããã ãã§ä¿¡é ŒãããŸã
ãããŠããã®4è¡ãæžããŠãããšãå幎åŸã«æ°åãå€ãã£ãŠããããªãã®å€æãééã£ãŠããããšã«ã¯ãªããŸãããåœæã®æåã ã£ããšèª¬æã§ããããã§ãã
ãã®èšäºã¯ãå®ç¹èŠ³æž¬ã«ããŸã
å幎ã§ãããã ãåæãå€ãããŸããã
æ³äººåããŠã幎å150ååã«ãªã£ãŠããã©ã³ãåãå€ãã£ãŠãåç»ãšãšãŒãžã§ã³ãã®ããŒããå¢ããŠãæ¹æ³è«ã«å¯Ÿããæ»èªä»ãã®æ¹å€ãåºãŸããã
ãããŠå幎åŸã«ã¯ããŸãå€ãã£ãŠããã¯ãã§ã
ãªã®ã§ããã®èšäºã¯å®æçã«æŽæ°çãåºã圢ã«ããŸãã
æŽæ°ãåãåãããæ¹ã¯ããã²ãã©ããŒããŠãããŠãã ãããæ°ããçãåºãããšãã«éç¥ãå±ããŸãã
ãããŠãArenaãèŠããšãã¯ãarena.ai ãããã¯ããŒã¯ããã ãã§ãªããèªåã®æ¥åã§ã©ã®ããŒããèŠãã¹ããã1ã€æ±ºããŠãããŠãã ããç·åãèŠç¶ãããããã倿ã«ã¯äœ¿ããŸããã
次ã«èªããã®
ã©ã³ãã³ã°ãä¿¡ããããªããªã£ãããšãå®åã§å¿ èŠã«ãªãã®ã¯ãèªç€Ÿã®ãŠãŒã¹ã±ãŒã¹ã§èªåã§æž¬ãæé ã§ãã
çšè°ã«æžããAIã¢ãã«éžå®
20åã®è©äŸ¡ã»ãããäœã£ãŠã3ã€ã®ã¢ãã«ã30åã§æ¯èŒããŠãçšè°æžã«æžãã圢ã«ãããšãããŸã§ããŸãšããŸããã
瀟å RAGã®ãã¢ã¯å®ç§ã ã£ããçŸå Žã«åºããã3ã¶æã§æ¢ãŸã£ãçç±
ç€Ÿå ææžãåç §ãããAIãæ€èšããŠããæ¹ã¯ããã¡ããã©ãããåãèãæ¹ããRAGã®è©äŸ¡ã«åœãŠã¯ããèšäºã§ãã
äžæ¬¡æ å ±
Arena, the AI leaderboard everyone uses, is now a $100M businessïŒTechCrunchã2026幎6æ29æ¥ïŒ
The Leaderboard IllusionïŒSingh et al., arXiv:2504.20879 / NeurIPS 2025 Datasets and Benchmarks TrackïŒ
Arena å ¬åŒ
èšèŒã¯2026幎8ææç¹ã®å ¬éæ å ±ã«ããšã¥ããŸãããªãŒããŒããŒãã®æ°å€ãšéå¶æ¹éã¯é »ç¹ã«æŽæ°ããããããå®éã®åç §æã«ã¯å¿ ãå ¬åŒãµã€ãã§ææ°ã®ç¶æ ã確èªããŠãã ããã
ãããããã¬ãžã³
çè X
AIææ°æ å ±(ããšã å€ã)ãçºä¿¡äžïŒ