
LLMã®äŸ¡æ Œã¯æ¬åœã«åŽ©ããŠããââäžãã£ãŠããã®ã¯ãããŒã¯ã³å䟡ãã§ã¯ãªããç¥èœã®åŒã³åºãäŸ¡æ Œãã
Stanford AI Index 2025ã«ããã°ãAIã®ãã³ãããŒã¯ã§ããMMLUã«ãããŠãGPT-3.5çžåœã®æ§èœãåºãããã®åãåããã³ã¹ãã¯ã2022幎11æã®100äžããŒã¯ã³åœãã20ãã«ããã2024幎10æã«ã¯0.07ãã«ãŸã§äžãã£ãã

ããã¯åãã¢ãã«ã®åçŽãªå€äžãã§ã¯ãªããäžå®æ°Žæºã®ãè³¢ãããåŒã³åºãããã®äŸ¡æ ŒããçŽ18ã«æã§å€§ããçµã¿æ¿ããã€ã€ããããšãæå³ããŠããã
å°ãåãŸã§ãçæAIã®APIå©çšæã¯ã倧éåŠçãå šç€Ÿå°å ¥ãèããæ®µã«ãªããšç¡èŠãã«ããå¶çŽã ã£ãããšãããä»ããã®åæã¯ãã§ã«å€§ããå€ãã£ãã
ãã ããçŸåšèµ·ããŠããã®ã¯ãAIãšããååšãã®ãã®ãå®ããªã£ãããšããåçŽãªçŸè±¡ã§ã¯ãªããåçã«å®ããªã£ãŠããã®ã¯ããããŸã§ãäžå®æ°Žæºã®ç¥èœãåŒã³åºãã³ã¹ããã§ããã
æ¬çš¿ã§ã¯ãAIã®äŸ¡æ Œäœäžã«ã€ããŠãç
œãã§ã¯ãªãæ§é çãªå€åãšããŠè§£ãæãããŠãããLLMã®æšè«ã³ã¹ãäœäžã®è£ã§äœãèµ·ããŠããã®ãããªãããã¯åå°äœã®é²åã ãã§ã¯èª¬æã§ããªãã®ãããããŠãããŒã¯ã³å䟡ã®äœäžãæå³ããããšãšãæå³ããªãããšã®å¢çç·ãåŒãããšããå®åã®çŸå Žã§ã¯ã©ã®ãããªã課é軞ã®ã·ããããèµ·ããŠããã®ãããããã®åããçŽè§£ãããšã§ãããããå
ã®AIããžãã¹ã«ãããŠäœã䟡å€ãšããŠæ®ãã®ããèŠéããŠã¿ããã
第1ç¯ æ¬åœã«æ¥èœããŠããã®ãââåºå®æ§èœæ¯èŒã瀺ããæšè«äŸ¡æ Œã®äœäžã
ãLLMã®æšè«ã³ã¹ããäžãã£ãŠããããšãã話é¡ã¯ãçæAIçéã§ã¯æ¥åžžçã«èªãããŠãããããããæ°ããã¢ãã«ãçºè¡šããããã³ã«æç€ºãããã以åã®ã¢ãã«ãããïŒ å®ãããšããæççãªæ å ±ã ãã远ã£ãŠããŠããå šäœåã¯èŠããŠããªããéèŠãªã®ã¯ãåãªãäŸ¡æ Œè¡šã®æ¯èŒã§ã¯ãªãããåçã®æ§èœãåŒãåºãããã«ããããããã®ãããšããåºå®æ§èœæ¯èŒã®èŠç¹ã§ããã
Stanford AI Index 2025ã®å ±åãèŠãã°ããã®å€åã®å€§ãããæ°å€ãšããŠæµ®ã圫ãã«ãªããåè¿°ã®éããMMLUã«ãããŠGPT-3.5çžåœã®æ§èœã¹ã³ã¢ãåºãããã«å¿ èŠãªAPIã³ã¹ãã¯ãçŽ18ã«æã§20ãã«ãã0.07ãã«ãžãšäœäžããŠããã
éå»ã«ã¯é«äŸ¡ã ã£ãäžå®æ°Žæºã®èœåããããŸã¯ããå®äŸ¡ã«åŒã³åºããããã«ãªã£ãŠãããéèŠãªã®ã¯ãåãååã®åçŽãªå€äžããšãããããéå»ã®é«æ§èœãæ®åäŸ¡æ Œåž¯ã«èœã¡ãŠããç¹ã§ãããããã«åã¬ããŒãã¯ãMMLUã ãã§ãªããããé«åºŠãªå°éç¥èãè«ççæšè«ãæ±ããããGPQAãªã©ã®ãã³ãããŒã¯ã«ãããŠããäžå®æ§èœæ°Žæºãè¶ ããã¢ãã«ã®ã³ã¹ããæç¢ºã«äœäžããŠããããšã瀺ããŠããã
AIã®é²åé床ããã¬ã³ããå°éã«åæãããªãµãŒãæ©é¢ãEpoch AIãã®èª¿æ»ãããã®åŸåãè£ä»ããŠãããEpoch AIã®ããŒã¿ã«ããã°ãåºå®æ§èœã§èŠãLLMã®æšè«äŸ¡æ Œã¯ããããã2ã«æã§åæžããšããæ¥éãªããŒã¹ã§äœäžããŠãããå°ãªããšããããŒããŠã§ã¢åäœã®äŸ¡æ Œæ§èœæ¹åã ãã§ã¯èª¬æããããªãé床ã§ãåºå®æ§èœãããã®æšè«äŸ¡æ Œãäžãã£ãŠããã®ã ã
ãã³ãããŒã¯æ¯èŒã®éçïŒå®å䟡å€ãšã®ä¹é¢
ããã§äžã€ãéèŠãªæ³šæç¹ãæç€ºããŠãããããããã¯ããããã§ããåçã¯ããããŸã§ãã³ãããŒã¯äžã®æ¯èŒã§ããããšããããšã ã
MMLUãGPQAã§éå»ã®ã¢ãã«ãšåãã¹ã³ã¢ãå©ãåºãããããšãã£ãŠãå®éã®ãããã¯ãã«çµã¿èŸŒãã éã®ããã³ãããžã®å¿ å®ãã埮åŠãªãã¥ã¢ã³ã¹ã®æ±²ã¿åãæ¹ããããã¯ç¹å®ã®å°éãã¡ã€ã³ã«ãããå®å®æ§ãªã©ãå®åã«ããã䟡å€ãå®å šã«äžèŽãããšã¯éããªãããã³ãããŒã¯ã®ã¹ã³ã¢ã¯ãããŸã§äžã€ã®ææšã§ãããçŸå Žã®èæèŠãšããŠã®ãè³¢ããã100%代åŒãããã®ã§ã¯ãªãã
ãããããã®éçãå·®ãåŒãããšããŠãã倧ããªæ¹åæ§ãæºããããšã¯ãªããå°ãªããšããäžå®æ°Žæºã®æ§èœãåŸãããã®æšè«ã³ã¹ãã倧ããäœäžããŠããããšèªäœã¯ãããªãå ã芳枬ã§ããããããŸã§ã³ã¹ãã®å£ã«é»ãŸããŠæãåºããªãã£ã倧éã®ããŒã¿åŠçããå šãŠãŒã¶ãŒåãã®ã¢ã·ã¹ã¿ã³ãå®è£ ãªã©ããçŸå®çãªäºç®å ã§å®è¡å¯èœã«ãªã£ãŠããã®ã ã
第2ç¯ ãªããããŸã§äžãã£ãã®ãââäŸ¡æ Œäœäžã®æ¬äœã¯åå°äœã ãã§ã¯ãªã
ã§ã¯ããªããããŸã§æ¥éã«æšè«ã³ã¹ããäžãã£ãŠããã®ã ããããæãçŽæçãªèª¬æã¯ãAIãåããããã®GPUãªã©ã®ããŒããŠã§ã¢ã鲿©ããããããšãããã®ã ããã確ãã«ããŒããŠã§ã¢ã®æè¡é©æ°ã¯ç¶ããŠãããããããããŒã¿ãå·éã«èªã¿è§£ããšãããã ãã§ã¯ãã®ã³ã¹ãäœäžã説æããããªãã
Stanford AI Index 2025ã«ããã°ãæ©æ¢°åŠç¿åãããŒããŠã§ã¢ã®äŸ¡æ Œæ§èœã¯ã幎ç30%ã®æ¹åãããããŠãšãã«ã®ãŒå¹çã¯ã幎ç40%ã®æ¹åããšæšå®ãããŠãããããã¯åªããæ°åã§ããããå ã»ã©èŠããæšè«äŸ¡æ Œã2ã«æã§åæžããããšããEpoch AIã®ããŒã¿ãšç §ããåãããŠã¿ããšã幎ç30ã40%ã®æ¹åããŒã¹ã§ã¯ãçŽ18ã«æã§280å以äžãäœäžããé床ã説æããããªãã
æšè«äŸ¡æ Œã®äœäžã®èæ¯ã«ã¯ãããŒããŠã§ã¢ã®æ¹åã«å ããŠããœãããŠã§ã¢ã®é²åããããã€ããŒåŽã®æäŸèšèšã®å·¥å€«ããããŠç«¶äºç°å¢ã®å€åãéãªã£ãŠãããšã¿ãã®ãèªç¶ã ã
第äžã«ãæè¡çãªæé©åãå¯äžãããšã¿ãããã巚倧ãªã¢ãã«ããã®ãŸãŸåããã®ã§ã¯ãªããæ§èœãç¶æãããŸãŸãµã€ãºãå°ãããããèžçïŒDistillationïŒãã®æè¡ããæšè«ããã»ã¹ãå¹ççã«æãããã®ãµãŒãã³ã°æé©åãé²ãã ã
第äºã«ãAPIãã©ãããã©ãŒããŒåŽã®ååèšèšïŒæäŸã®ä»æ¹ïŒã®å·¥å€«ã§ãããäž»èŠãã¬ã€ã€ãŒã¯ãããããåã«ã¢ãã«ã®äŸ¡æ Œãäžããã ãã§ãªããã·ã¹ãã ã®ç©ºããªãœãŒã¹ãæå¹æŽ»çšããä»çµã¿ãå°å ¥ãå§ããŠããã峿æ§ãæ±ããããªãã¿ã¹ã¯ããŸãšããŠå®äŸ¡ã«åŠçããããããïŒBatchïŒåŠçãããäžåºŠèšç®ããæèãäžæçã«ä¿åããŠåå©çšããããã£ãã·ã¥ïŒPrompt CachingïŒããšãã£ãæ©èœãããã ã
第äžã«ãåžå Žã«ãããäŸ¡æ Œç«¶äºã®åŽé¢ã§ããã2026幎2æãReutersã¯ãDeepSeek以éã®äœäŸ¡æ Œã¢ãã«ç«¶äºãæ¥çå šäœã®äŸ¡æ Œå§åãé«ããããšå ±ããŠãããçŸåšã®APIäŸ¡æ Œè¡šã¯ãçŽç²ãªå䟡衚ã§ã¯ãªããã·ã§ã¢ç²åŸããšã³ã·ã¹ãã æ§ç¯ã®æŠç¥ãå«ãã ãåžå ŽäŸ¡æ ŒãšããŠèªãã¹ãé¢ããããããããè€åçã«çµ¡ã¿åãããšã§ãæšè«ã³ã¹ãã¯æŒãäžããããŠããã

第3ç¯ ããã§ããAIãå®ããªã£ãããšèšãåããªãçç±ââèšç·Žã³ã¹ããšå®åã®ç·é¡
ãããŸã§æšè«äŸ¡æ Œã®äœäžãèŠãŠãããšããAIå šäœãäžæ§ã«å®ããªã£ãããšé¯èŠããŠããŸããããããªããããããæšè«ã³ã¹ããäžãã£ãŠããããšãšããAIå šäœãå®ããªã£ãããšãã¯å¥ã®è©±ã ã
ãŸããæ°ããªãç¥èœããçã¿åºãããã®æºæ³ã§ãããèšç·Žã³ã¹ããã¯ãäŸç¶ãšããŠéããAI Index 2025ã«ãããã¬ãŒãã³ã°ã³ã³ãã¥ãŒãã³ã¹ãã®æšå®ã§ã¯ãGPT-4ã¯çŽ7900äžãã«ãLlama 3.1 405Bã¯çŽ1.7åãã«ã«éãããæå 端ã®ããã³ãã£ã¢ã¢ãã«ãéçºããããã«ã¯ãããŒã¿åéãè©äŸ¡ãçµç¹ã®äººä»¶è²»ãªã©ãå«ããªãçŽç²ãªèšç®è³æºã ãã§ããèšå€§ãªæè³ãå¿ èŠãªãŸãŸã ã

ããã«éèŠãªã®ã¯ãå®åã®çŸå Žã«ãããã³ã¹ãæ§é ã®å€åã§ããã
APIã®ããŒã¯ã³å䟡ãäžãã£ããããšãã£ãŠãå®éã®è«æ±é¡ãåçã«äžãããšã¯éããªããå䟡ãå®ããªã£ãããšã§ãç§ãã¡ã¯ããé«åºŠã§è€éãªã¿ã¹ã¯ãAIã«ä»»ããããã«ãªã£ãŠããããã ã
ããšãã°ãçãããã¹ãã®èŠçŽ1åãªãå®ããããããäœäžåãšããé·æè³æãèªã¿èŸŒã¿ãç€Ÿå ææžãæ€çŽ¢ããèªåŸçã«æèãç¹°ãè¿ãreasoningïŒæšè«ïŒãçµãŠãæ ¹æ ãæ·»ããäžæžããè¿ããããªã¯ãŒã¯ãããŒã«ãªããšãæ¶è²»ãããç·ããŒã¯ã³æ°ãå©çšæ©èœã¯å¢å€§ããåçŽãªå ¥å/åºåããŒã¯ã³å䟡ã ãã§ã¯ç·é¡ã説æã§ããªããªãã
ãããã€ããŒåŽãããã«åãããŠäŸ¡æ Œäœç³»ãå€åãããŠãããGoogleã¯Gemini APIã®äŸ¡æ Œèª¬æã«ãã¢ãã«ãå éšã§è«ççãªã¹ããããèžãã§ãèãããããã«æ¶è²»ããããthinking tokensïŒæèããŒã¯ã³ïŒãããäºå®ç¢ºèªã®ããã®ãgroundingããçµã¿èŸŒãã§ãããOpenAIã®APIã«ãããŠããWeb SearchãFile Searchãªã©ã®ããŒã«å©çšã«å¯ŸããŠã¯å¥å»ºãŠã®æéãèšå®ãããŠããã
ãããã®å€åãèžãŸãããšãå°ãªããšãå®åã®çŸå Žã§ã¯ãã³ã¹ããèŠãåäœãã1MããŒã¯ã³ãããããããã1ä»¶ã®ä»äºãçµããã®ã«ããããããžç§»ãã€ã€ããããšèããã»ãã宿 ã«è¿ãã

第4ç¯ äœãã³ã¢ãã£ãã£åããäœãæ®ãã®ãââãä»äºã®å®äºããå·¡ãæ°ããªäŸ¡å€
å°ãªããšãå®åã§ã¯ãã³ã¹ããèŠãåäœãä»äºå®äºãžå¯ãã€ã€ããããã®ãšããããžãã¹ã«ãããŠäœãã³ã¢ãã£ãã£åããäœãæ°ããªäŸ¡å€ãšããŠæ®ãã®ã ãããã
ãŸãã³ã¢ãã£ãã£åããããã®ã¯ãåçºã®èŠçŽãåé¡ãšãã£ãã1åã®APIã³ãŒã«ã§è¿ããå®åçãªåŠçã§ããããããã¯ã©ã®ã¢ãã«ã䜿ã£ãŠãå®äŸ¡ã«å®çŸã§ããã
äžæ¹ã§äŸ¡å€ããããé«ãŸã£ãŠããã®ã¯ããšã³ãžãã¢ãªã³ã°ãšæ¥åèšèšã®é åã§ãããé»åã®å䟡ãäžãã£ãŠããå·¥å Žã®ææ¥ã³ã¹ããåãå²åã§äžãããšã¯éããªãããã«ãå·®ãã€ããããã®ã¯ãåå¥ã®ããã³ãããã®ãã®ãããå®ããªã£ãç¥èœãã©ãçµã¿åãããŠå®åã«èããæµããäœãããšããèšèšã®åŽã«ãªãã€ã€ããã
䟡å€ãæ®ãã®ã¯ãå®ããªã£ãã¢ãã«ãçµã¿åãããŠä»äºã®æµããèšèšãããã®åºåãè©äŸ¡ããéçšã®äžã§ã³ã¹ããšå質ãäž¡ç«ãããèœåã§ããã

å ·äœçã«ã¯ãã¯ãŒã¯ãããŒèšèšãã ãå°åã¢ãã«ãšå€§åã¢ãã«ãã©ã䜿ãåãããããŠãŒã¶ãŒã®å ¥åãã©ãæ¯ãåããããæ€çŽ¢ïŒRAGïŒãå€éšããŒã«çµ±åãã©ãçµã¿åãããããããšãã°ãæåã®åãåããåé¡ã¯å®äŸ¡ãªå°åã¢ãã«ã§è¡ããè€éãªèŠä»¶ã®ã¿ã倧åã¢ãã«ã«åããããã«ç€Ÿå ææžæ€çŽ¢ãæãã§æ ¹æ ä»ãã®åçãè¿ããšãã£ãèšèšã§ãããããããè€æ°ã®ã³ã³ããŒãã³ãã飿ºãããèšèšå³ããã·ã¹ãã ã®æçµçãªããã©ãŒãã³ã¹ãšã³ã¹ãå¹çãæ±ºå®ã¥ããã

ãããŠãã¯ãŒã¯ãããŒãæŽç·Žããè€éãªã¿ã¹ã¯ãAIã«ä»»ããããã«ãªãã»ã©ãä»åºŠã¯åºåãè©äŸ¡ããä¿¡é Œæ§ãæ ä¿ããä»çµã¿ããäžå¯æ¬ ã«ãªããAIãéäžã§ééã£ãæšè«ãããŠããªãããç£èŠããè©äŸ¡ããããã»ã¹ãã©ãçµã¿èŸŒãããåãããã
ããã«ããã®è©äŸ¡ããã»ã¹ãå«ããã·ã¹ãã å šäœãåãç¶ãããéçšèšèšããéèŠã ããã£ãã·ã¥ããããåŠçãšãã£ãæäŸåœ¢æ ã®å·¥å€«ãå·§ã¿ã«äœ¿ãããªãããŠãŒã¶ãŒäœéšãæãªãããšãªãã³ã¹ãããã©ãŒãã³ã¹ãæå€§åããã€ã³ãã©çèŠç¹ãæ±ããããã

çµè«
ããŒã¯ã³å䟡ã®äœäžã¯ãçæAIã®æ®åãé»ãã§ããå£ã確ãã«äœããããã ãããã®ããšã¯ãAIå šäœãå®ããªã£ããããšããã®ãŸãŸæå³ããªãããããèµ·ããŠããã®ã¯ãäžå®æ°Žæºã®ç¥èœãæ¥éã«ã€ã³ãã©åãããã®äžã§ãã©ã®ä»äºããã©ã®èšèšã§ãã©ããŸã§èªååããããã®äŸ¡å€ãåé¢ã«åºãŠããããšããæ§é å€åã§ããã
å®ããªã£ãã®ã¯ãç¥èœãã®ãã®ãšããããããã®åŒã³åºãã®æäœã³ã¹ãã§ãããå®åã®çŠç¹ã¯ãã§ã«ããã®ç¥èœãã©ãçµã¿èŸŒã¿ãã©ããŸã§å®äºè²¬ä»»ãæããããããžãšç§»ãã€ã€ããã
ç§ãã¡ãåãåãã¹ãã¯ãåãªãã³ã¹ãããŠã³ã®è°è«ã§ã¯ãªããã€ã³ãã©åããç¥èœãé§äœ¿ããŠããã«é«ä»å 䟡å€ãªä»äºãå®äºãããããšããæ ¹æ¬çãªåããªã®ã§ããã

ãæ³šéã
èšäºäžã®èšç·Žã³ã¹ãïŒGPT-4 çŽ7900äžãã«ãLlama 3.1 405B çŽ1.7åãã«ïŒã¯ãStanford AI Index 2025ã«ããããã¬ãŒãã³ã°ã³ã³ãã¥ãŒãã³ã¹ãã®æšå®å€ã«åºã¥ããŸãã
åè
æ¬çš¿ã®å·çã«ãããã以äžã®ã¬ããŒãããã³å
Œξ
å ±ãåç
§ã»åºç€ãšããŠããïŒâ»æ
å ±ã¯ãã¹ãŠåè³æã®å
¬éæç¹ããŸãã¯çŸè¡ã®ä»æ§ã«åºã¥ãïŒã
Stanford AI Index 2025
åºå®æ§èœã§èŠãæšè«ã³ã¹ãã®äœäžïŒMMLUã§ã®æ¯èŒïŒãæ©æ¢°åŠç¿åãããŒããŠã§ã¢ã®äŸ¡æ Œæ§èœïŒå¹Ž30%æ¹åïŒããã³ãšãã«ã®ãŒå¹çïŒå¹Ž40%æ¹åïŒã®æšç§»ããªãã³ã«GPT-4ãLlama 3.1 405Bã®èšç·Žã³ã¹ãæšå®å€ãªã©ãäž»èŠããŒã¿ã®åºç€ãšããŠããã
https://hai-production.s3.amazonaws.com/files/hai_ai_index_report_2025.pdfEpoch AI trends
åºå®æ§èœã§èŠãLLMæšè«äŸ¡æ Œã¯ãããã2ã«æã§åæžããšããã³ã¹ãäœäžã®é床æã®è£ä»ããšããŠåç §ã
https://epoch.ai/trendsOpenAI API Pricing
çŸè¡ã®äŸ¡æ Œäœç³»ãBatchåŠçãWeb SearchãFile Searchçã®å¥å»ºãŠèª²éæ§é ã®ç¢ºèªã
https://openai.com/api/pricing/Google Gemini API Pricing
ã¢ãã«å¥äŸ¡æ Œã«å ãããthinking tokensããcacheãgroundingãªã©ãä»äºåäœã³ã¹ããžã®ç§»è¡ã瀺ã仿§ã確èªã
https://ai.google.dev/gemini-api/docs/pricingAnthropic Claude Pricing
ã¢ãã«äŸ¡æ Œã®æ§é ãšãbatch processingãªã©ã®å©çšåœ¢æ ã«ã€ããŠç¢ºèªã
https://platform.claude.com/docs/en/about-claude/pricingReutersïŒ2026幎2æ12æ¥ä»å ±éïŒ
DeepSeek以éã®äœäŸ¡æ Œã¢ãã«ç«¶äºãäŸ¡æ Œå§åã匷ãããšããåžå Žååã®ç¢ºèªã
https://www.reuters.com/world/china/year-deepseek-shock-get-set-flurry-low-cost-chinese-ai-models-2026-02-12/