Izinguquko zeClaude Opus 4.7 kubathuthukisi: funda ngokujulile imodeli entsha

Ikhaya » Python » Izinguquko zeClaude Opus 4.7 kubathuthukisi: funda ngokujulile imodeli entsha

I-Claude Opus 4.7 isetshenziswa njengemodeli entsha ye-Anthropic yokutholakala okujwayelekile, futhi kubathuthukisi izwakala ingafani nhlobo nokushaqeka okuncane kodwa ifana kakhulu nesizukulwane esisha samathuluzi. Ukugxila kwalokhu kukhishwa kucacile: imisebenzi enzima yobunjiniyela besofthiwe, ukuhamba komsebenzi okude okuzimela, ukuqonda okujulile kwe-multimodal, kanye nokulawula okuqinile kokuthi imodeli "icabanga" futhi ichitha malini. Uma ubulokhu ucindezela izinguqulo zangaphambilini ze-Opus emikhawulweni yazo ngama-ejenti okubhala ikhodi ayinkimbinkimbi, umsebenzi wolwazi oluningi, noma ukuzenzekela okusekelwe embonweni, lesi sibuyekezo sishintsha umdlalo ngezindlela eziningana ezisebenzayo.

Okugqama kakhulu ukuthi i-Anthropic ayigcinanga nje ngokuthuthukisa ikhono layo lokukhiqiza; iphinde yashintsha izindlela zayo zokuziphatha ezibalulekile, Ama-API kanye nezithiyo zokuvikela ngakho-ke i-Opus 4.7 iziphatha njengomlingani weqembu onokwethenjelwa ongashiywa esebenza isikhathi eside. Ukulandela imiyalelo kuqinile, inkumbulo isebenziseka kalula kumaphrojekthi angempela, ukwesekwa kwe-multimodal kukhuphukela kusisombululo esiphezulu sangempela, futhi ipulatifomu ezungezile (i-Claude Code, isabelomali semisebenzi, amazinga omzamo) ilungisiwe ngokuphathelene nokusebenza konjiniyela bomhlaba wangempela. Lokho kusho ukuthi kuzodingeka uphinde ubuyekeze ezinye izikhuthazo kanye ne-infra, kodwa uma uzivumelanisa nezimo, ungalayisha okuningi "kwezinto ezinzima nezicasulayo" kunangaphambili.

Lokho uClaude Opus 4.7 azama ukuba yikho konjiniyela

I-Anthropic ibeka iClaude Opus 4.7 njengemodeli yayo enekhono kakhulu etholakalayo ngokuvamile, elungiselelwe ngqo umsebenzi "we-agent" osezingeni elide kanye nemisebenzi yolwazi oluthuthukisiwe. Ngamagama alula, yakhelwe ukusebenzisa isofthiwe yezinyathelo eziningi kanye nemisebenzi yedatha enezibambo ezimbalwa, ukuzihlola okungcono kanye nokuziphatha okuvumelanayo kakhudlwana ngezikhathi eziningi.

Uma kuqhathaniswa ne-Opus 4.6, imodeli entsha ikhombisa izinzuzo eziphawulekayo kubunjiniyela besofthiwe obuthuthukisiwe, ikakhulukazi emisebenzini eyinselele kakhulu eyayidinga ukugadwa okuseduze kwabantu. Abasebenzisi bokuqala babika ukuthi manje sebengakwazi ukudlulisela izinkinga zabo ezinzima kakhulu zokubhala ikhodi—kufaka phakathi ama-refactor agxilile, izinguquko zezinsizakalo eziningi kanye nokulungisa amaphutha okunzima—ku-Opus 4.7 ngokuzethemba okukhulu. Imodeli ayikhiqizi nje ikhodi; ihlela, isebenzise futhi iqinisekise kabusha umsebenzi wayo ngokuqinile.

Ngaphansi kwe-hood, i-Opus 4.7 igcina iwindi elifanayo lokuqukethwe kwe-1M-token kanye namathokheni okukhipha afinyelela ku-128k enza izinguqulo zangaphambilini ze-Opus zikhange kuma-codebase amakhulu namadokhumenti amakhulu. Iphinde isekele ukucabanga okuguquguqukayo, ithuluzi elijwayelekile kanye nesethi yesikhulumi esivela ku-Opus 4.6, futhi itholakala noma yikuphi lapho ungalindela khona: nge-Claude API, yonke imikhiqizo ye-Claude, i-Amazon Bedrock, i-Google Cloud Vertex AI kanye ne-Microsoft Foundry. Intengo ihlala ifana: cishe amathokheni okufaka ayizigidi ezi-5 kanye namathokheni okukhipha ayizigidi ezingama-25.

Ngale kwemingcele ye-raw throughput kanye ne-token, imodeli izimele ngamabomu endleleni ephatha ngayo ama-loop “afana ne-agent”. Ihlanganiswe nemibono emisha efana nesabelomali semisebenzi kanye nepharamitha yomzamo ocwengekile, i-Opus 4.7 ihloselwe ukufakwa ezinhlelweni lapho ingahlela khona izinyathelo eziningi—ukusetshenziswa kwamathuluzi, ukuhlela amafayela, ukuhlola, izifinyezo—ngaphandle kokubuyisela ukulawula kumsebenzisi njalo.

Ukulandela imiyalelo okubukhali kanye nokushintsha kokuziphatha okufanele abathuthukisi bazivumelanise nakho

Omunye wemibono emikhulu yansuku zonke abathuthukisi abazoyizwa ukuthi i-Opus 4.7 ilandela kanjani imiyalelo uma iqhathaniswa namamodeli amadala e-Claude. Lapho i-Opus 4.6 kanye nontanga ngezinye izikhathi beyifihla imininingwane, beyihlanganisa kusukela kwesinye isibonelo kuya kwesinye, noma beqa buthule izingxenye ezingacacile zesaziso, i-Opus 4.7 izama kanzima kakhulu ukuhambisana ngqo nalokho okubhalile—hhayi okunye okuningi.

Lokhu kulandela imiyalelo eqinile kunomphumela owusizo: izixwayiso ezazisebenza “kahle ngokwanele” ngaphambili manje sezingaqala ukuziphatha ngezindlela ezimangalisayo. Uma ama-harnesses akho amadala ayethembele kumodeli ukuze anqume izinyathelo ongazipelanga ngokugcwele, noma ukuze ahlanganise amaphethini kusukela entweni eyodwa yohlu kuya kwenye, manje ungase ubone umphumela ozwakala uqinile noma ungaphelele. I-Anthropic incoma ngokucacile ukuthi kubuyekezwe futhi kulungiswe kabusha izixwayiso kanye nama-harnesses okuhlanganisa ngenxa yalesi sizathu.

Ubude bempendulo bulinganiswa ngendlela ehlukile: I-Opus 4.7 ilungisa ukuphindaphindwa komsebenzi wayo obonakalayo ekubeni yinkimbinkimbi yomsebenzi esikhundleni sokushintsha isitayela sempendulo sobude obungaguquki. Ngemibuzo elula, uzothola izimpendulo ezimfushane kakhudlwana, kuyilapho imisebenzi enzima enezingxenye eziningi noma e-ejensi ikhiqiza ngokwemvelo imiphumela emide nenemininingwane eminingi. Lokhu kulinganisa okuguquguqukayo kusho ukuthi uzofuna ukucacisa kakhudlwana ngokufushane uma uhlelo lwakho lokusebenza lunemikhawulo eqinile yokukhipha.

Imodeli isebenzisa futhi izingcingo ezimbalwa zamathuluzi ngokuzenzakalelayo futhi incike kakhulu ekucabangeni kwayo, ngaphandle kokuthi uyidonsela ngamabomu ekusebenziseni ithuluzi ngokukhulisa izinga lomzamo. Lezo yizindaba ezinhle ngokubambezeleka kanye nezindleko ezimweni eziningi, kodwa uma uhlelo lwakho lulindele ukwakheka okunzima kakhulu ngamathuluzi (isb., ukusebenza kwekhodi okunolaka, ama-linters noma ama-simulators), kufanele uhlole ukuthi udinga yini ukukhuthaza ukusetshenziswa kwamathuluzi amaningi ngokukhuthaza kanye nokulungisa umzamo.

Ngokwezwi, i-Opus 4.7 iyashintsha isitayela se-Opus 4.6 esifudumele kakhulu, esiqinile futhi esiqinisekisayo iye ezwini eliqondile, elinemibono eminingi elinama-emoji ambalwa kanye nokuchuma okuncane kokuthi “wenza kahle”. Ngokusetshenziswa kwamathuluzi onjiniyela, lokho kuvame ukuba yimpumelelo: uthola izahlulelo ezicacile, ukugxekwa okuqondile kanye nokugqoka okuncane. Kuzinhlelo zokusebenza zabasebenzisi ezithembele ekubhalweni okulula, ungase ufune ukuqondisa ithoni ngokucacile ezicelweni zakho.

Umbono ophakeme kanye nezinzuzo eziningi ezibalulekile kumaphrojekthi angempela

I-Opus 4.7 iyimodeli yokuqala ye-Claude enokusekelwa kwesithombe okunesinqumo esiphezulu, okwandisa usayizi wesithombe esiphezulu sibe amaphikseli angu-2576 onqenqemeni olude, noma cishe ama-megapixel angu-3.75. Lokho kungaphezu kokuphindwe kathathu kunomkhawulo wangaphambilini we-1568px / 1.15MP, futhi kushintsha kakhulu lokho ongakunikeza imodeli ngokuphephile lapho usebenzisana nezinto ezibonakalayo ezixineneyo.

Lokhu kukhuphuka kwesisombululo kuvula ngqo ukusebenza okungcono emisebenzini enzima yokubona: ama-ejenti manje angahlaziya ngokuthembekile izithombe-skrini ze-UI ezigcwele abantu, imidwebo eyinkimbinkimbi, kanye nokuskena kwamadokhumenti okuqukethe imininingwane ecacile. Sebenzisa izimo ezifana nama-ejenti okusebenzisa ikhompyutha ufunda izithombe-skrini ezigcwele zedeskithophu, ukudonsa idatha ehlelekile kusuka kumashadi ayinkimbinkimbi, noma ukuqhathanisa ukwakheka okuphelele kwamaphikseli kuba ngcono kakhulu ngaphandle kokunciphisa konke kuze kube sezingeni lokulahlekelwa ulwazi.

I-Anthropic iphinde yenza kube lula ukuphatha ama-coordinate ukuze ama-coordinate angaphakathi emodeli ahambisane no-1:1 nama-pixel angempela esithombeni. Lokho kusho ukuthi uma uhlela amabhokisi omngcele, uchofoza ama-targets noma ubeka izichasiselo ngaphezulu, akudingeki ukuthi ulinganisele izici zakho zesikali. Kulula kakhulu ukuthi “chofoza ku-(x, y)” ngokusekelwe kumphumela wemodeli futhi uthembe ukuthi ihambisana kahle nesithombe osithumele.

Ngale kwesisombululo esimsulwa, i-Opus 4.7 ithuthukisa ukuqonda okuphansi kanye nokwenziwa kwendawo: ingcono ekubhekiseni, ekulinganiseni, ekubaleni kanye nemisebenzi efanayo ye-granular, futhi ukutholwa kwayo kwebhokisi lokulinganisa isithombe semvelo kunembe kakhulu. Kwabathuthukisi abakha ama-ejenti okuhlola i-UI, amapayipi e-QA abonakalayo, noma ama-bots okuhlaziya ishadi, lokhu kulungiswa okuzwakala kuncane kuhumusha kube amaphutha ambalwa ekutholeni izibalo ezihambisanayo kanye nezinto.

Yiqiniso, kukhona ukuphikisana: izithombe ezinesinqumo esiphezulu zidla amathokheni amaningi. Uma ungadingi ukwethembeka okwengeziwe, i-Anthropic iphakamisa ukuthi wehlise amasampula ezithombe ngaphambi kokuzithumela kumodeli ukuze kulawulwe ukusetshenziswa kwamathokheni. Kodwa uma udinga iphikseli ngayinye—isibonelo, lapho ubhala idatha yeshadi ezingeni lephikseli noma uqinisekisa ukwakheka kwamaslayidi kuye kumalebula ngamanye—imikhawulo emisha iwukunqoba okucacile.

Umsebenzi wolwazi, ezezimali kanye nemisebenzi yobuchwepheshe yamadokhumenti

I-Opus 4.7 akuyona nje intuthuko yokubhala amakhodi; iphinde ithumele amaphuzu aqinile ezilinganisweni zomsebenzi wolwazi, ikakhulukazi ezindaweni ezifana nezezimali kanye nomthetho lapho ukucabanga ngokunemba kanye nokubhala imibhalo ehlukahlukene kubaluleke kakhulu. Ekuhlolweni kwe-Anthropic's Finance Agent yangaphakathi, i-Opus 4.7 ifinyelela ukusebenza kwesimanje futhi isebenza njengomhlaziyi onekhono omncane kunomkhiqizi wombhalo olula.

Ekuhlolweni kwangaphakathi, imodeli ikhiqize amamodeli ezezimali aqinile kanye nokuhlaziywa kune-Opus 4.6, enezindaba ezihlelwe kangcono kanye nokuqagela okucacile. Futhi yenze umsebenzi ongcono wokuhlanganisa imisebenzi emincane eminingi—ukuqoqwa kwedatha, ukumodela ngezinombolo, ukwakha isethulo—kube yimiphumela ehambisanayo kusukela ekuqaleni kuya ekugcineni ebukeka futhi izwakala iwubungcweti.

I-Opus 4.7 iphinde ifinyelele imiphumela yesimanje ku-GDPval-AA, okuyisilinganiso senkampani yangaphandle esigxile emsebenzini wolwazi olubalulekile kwezomnotho kuzo zonke izizinda zezimali, ezomthetho kanye nezihlobene. Lokho kusho ukuthi lokhu akuzona nje izinzuzo zangaphakathi ezikhethwe kahle: imodeli isebenza kahle kakhulu kuneyangaphambili ezinkingeni eziyinkimbinkimbi, zokucabanga ezisetshenziswayo lapho imali yangempela noma ingozi isengozini.

Ngasohlangothini lwedokhumenti, i-Opus 4.7 ingcono kakhulu emisebenzini yokusebenza ehilela ukukhiqiza bese ihlola amafayela ehhovisi njenge-.docx kanye ne-.pptx. Kuthuthuke kakhulu ekukhiqizeni izinguquko ezilandelelwe kumafayela e-Word, ukulungisa izakhiwo ku-PowerPoint, bese kufundwa kabusha lezo ziphumelo—ngamathuluzi noma umbono—ukuqinisekisa ukuthi imiklamo yamaslayidi kanye nezimpawu zilungile. Uma izixwayiso zakho ngaphambilini kwadingeka zichaze ngokweqile ukuthi “hlola kabili isakhiwo samaslayidi ngaphambi kokusibuyisela,” ungase ukwazi ukususa ezinye zalezo zisekelo.

Amashadi kanye nokuhlaziywa kwezibalo nakho kuyazuza emandleni amasha e-multimodal. I-Opus 4.7 ingcono ekubizeni amathuluzi angaphandle njengemitapo yolwazi ye-Python imaging (isb., i-PIL) ukuhlola amashadi, ukukhipha idatha yezinga lamaphikseli nokuhumusha lezo zithombe zibe amasethi edatha noma izincazelo ezihlelekile. Leyo nhlanganisela—ukushayela ithuluzi kanye nombono obukhali—yenza isetshenziswe kakhulu njengomlingani kumadeshibhodi okuhlaziya kanye nemibhobho yokubika.

Imemori, ama-ejenti aqonde phezulu kanye nama-scratchpad esistimu yamafayela

Enye indawo lapho i-Opus 4.7 ithuthuka khona buthule kodwa ngendlela ebonakalayo yinkumbulo, ikakhulukazi ekusethweni lapho umenzeli angabhalela futhi afunde kusuka kufayela lamanothi aqhubekayo noma esitolo esihlelekile. Esikhundleni sokuphatha isicelo ngasinye njengesiqalo esisha kakhulu, imodeli ingcono ekunqumeni ukuthi yimiphi imininingwane okufanele ibhalwe phansi, ukuthi ibhalwe kanjani, nokuthi izophinde isetshenziswe nini esikhathini esizayo. tolerancia a fallos en búsqueda distribuida en flujos largos que dependen de contexto persistente.

Uma i-ejenti yakho igcina i-scratchpad, idokhumenti yamanothi noma isizindalwazi sememori esilula phakathi kokujika, i-Opus 4.7 kufanele ibonise ukuthuthuka okucacile ekhonweni layo lokusebenzisa lowo mongo wangaphandle. Kuzokhumbula njalo izinqumo ezibalulekile zephrojekthi, imiphumela engaphelele, kanye nezinto okufanele zenziwe emisebenzini eminingi, kunciphisa inani okufanele uliphinde kukho konke okushiwoyo.

U-Anthropic usho ngokusobala ukuthi uClaude manje usengcono ekubhaleni nasekusebenziseni inkumbulo esekelwa yisistimu yamafayela, okuwusizo kakhulu kuma-ejenti okubhala amakhodi ayinkimbinkimbi noma ocwaningo. Isibonelo, uma une-bot yokulungisa kabusha ezimele elandelela izinkinga ezivulekile, izinqumo zokwakha kanye nokuhlolwa okulindile ngaphakathi kweqoqo lamafayela, i-Opus 4.7 ngokuvamile izohlela futhi ihlole lolo lwazi ngokucabangela kakhulu kune-Opus 4.6.

Uma ungafuni ukugoqa ungqimba lwakho lwenkumbulo, i-Anthropic inikeza ithuluzi lememori eliseceleni kweklayenti elisebenza njenge-scratchpad ephethwe yiClaude. Lokhu kukuvumela ukuthi uzame ama-ejenti ahlala isikhathi eside—izikhathi ezihlanganisa amaseshini, amagatsha noma ngisho amasonto omsebenzi—ngaphandle kokwakha kuqala isizindalwazi se-vector esigcwele noma isevisi yamanothi enziwe ngokwezifiso.

Kuma-trace asebenza kahle afinyelela phezulu, imodeli ivame ukunikeza umsebenzisi izibuyekezo zenqubekela phambili ezivamile. Lokho kusho ukuthi uma ngaphambilini wengeze izixwayiso ze-scaffolding eziyinkimbinkimbi ukuze nje uphoqe imiyalezo "yesimo" ngezikhathi ezithile, ungazama ukwenza lula nokuvumela ukuziphatha okuzenzakalelayo kwe-Opus 4.7 ukuthi kuphathe ukubika inqubekela phambili, ikakhulukazi emazingeni aphezulu omzamo ku-Claude Code.

Ukuphepha, ukuhambisana kanye nokuphepha kwe-inthanethi

Ngokuphathelene nokuphepha nokuqondanisa, ukuhlolwa kwe-Anthropic kukhombisa ukuthi i-Opus 4.7 inephrofayili yengozi efanayo ne-Opus 4.6, enamazinga aphansi okuziphatha angakhathaza iningi labathuthukisi: ukukhohlisa, ukuvumelanisa kanye nokubambisana nokusebenzisa kabi. Ngezindlela eziningana, empeleni kuphephile kancane.

Imodeli ithola amaphuzu angcono ngokwethembeka kanye nokumelana nokuhlaselwa okunonya okusheshayo, okubaluleke kakhulu uma wakha ama-ejenti adla okuqukethwe okungathembekile kuwebhu, i-imeyili noma amadokhumenti akhiqizwe ngabasebenzisi. Ukumelana okunamandla komjovo kwenza kube nzima ukuthi izinto ezifakwayo eziphikisanayo zintshontshe imiyalelo yomodeli noma zikhiphe izimfihlo ngamaqhinga ahlakaniphile okusiza ngokushesha.

Nokho, kunezindawo ezimbalwa lapho i-Opus 4.7 ibuthakathaka khona kancane kune-Opus 4.6—ngokwesibonelo, ukuthambekela kwayo kokunikeza iseluleko esiningiliziwe sokunciphisa umonakalo mayelana nezinto ezilawulwayo. I-Anthropic isaphetha ngokuthi imodeli "ihambisana kahle kakhulu futhi ithembekile, yize ingeyona ephelele ngokuphelele," futhi iphawula ukuthi iClaude Mythos Preview ehlolwayo kakhulu isalokhu iyimodeli yabo ehambisana kahle kakhulu ngezilinganiso zangaphakathi.

Lokhu kukhishwa yilapho i-Anthropic iqala khona ukusebenzisa izindlela zokuphepha zokuphepha kwe-inthanethi ngesikhathi sangempela kumodeli evamile, kulandela ucwaningo kanye nemiyalezo evela emsebenzini wabo we-Project Glasswing. I-Opus 4.7 ifaka phakathi izinhlelo ezithola futhi zivimbele ngokuzenzakalelayo izicelo ezihlobene nezihloko zokuphepha kwe-inthanethi ezingavunyelwe noma eziyingozi kakhulu, ikakhulukazi lapho inhloso ibukeka isolisa.

Okubalulekile, i-Anthropic yenza umehluko phakathi kokusetshenziswa kabi kanye nomsebenzi wokuphepha osemthethweni: uma unguchwepheshe wezokuphepha owenza ucwaningo lobungozi, ukuhlolwa kokungena noma ukuhlangana kweqembu elibomvu, ukhuthazwa ukuthi ufake isicelo sohlelo lwabo lokuqinisekisa i-Cyber. Lolu hlelo luhloselwe ukunikeza ochwepheshe abahlolwe ukufinyelela amakhono ahlobene nokuphepha kwe-inthanethi e-Opus 4.7 ngaphandle kokuvula umnyango wokusetshenziswa kabi kakhulu, futhi izifundo ezifundwe lapha zizoqondisa izinqumo zokugcina mayelana nokukhishwa kwamamodeli ekilasini le-Mythos kabanzi.

Izilawuli zonjiniyela abasha: amazinga omzamo, ukucabanga okuguquguqukayo kanye nesabelomali semisebenzi

I-Opus 4.7 yethula isethi "yezinkinobho" eziyinkimbinkimbi kakhulu zabathuthukisi ukuze bahwebe ngamakhono, isivinini kanye nezindleko, kanye nepharamitha yomzamo ephakathi nendawo. Umzamo ulawula ukuthi imodeli icabanga kangakanani ngaphambi kokuphendula nokuthi isebenzisa amathuluzi ngobudlova kangakanani, okuthinta ngqo ukubambezeleka kanye nokusetshenziswa kwethokheni.

Ushintsho oluyinhloko luyizinga elisha lomzamo ophakeme kakhulu, xhigh, eliphakathi kwe-high ne-max futhi manje seliyi-default ye-Claude Code kuzo zonke izinhlelo. Ezimweni zokusebenzisa ikhodi kanye ne-ejenti, i-Anthropic incoma ukuqala nge-high noma i-xhigh, igcinela i-max kuphela izinkinga ezinonya kakhulu. Umzamo ophezulu usho usesho olubanzi, ukucabanga okujulile kanye nokuthembeka okungcono ngokuvamile—kodwa futhi namathokheni amaningi okukhipha kanye nezikhathi zokusebenza ezinde.

I-Opus 4.7 isusa ngokuphelele umqondo omdala "wesabelomali sokucabanga esinwetshiwe". Uma uzama ukusetha ukucabanga: {“type”: “enabled”, “budget_token”: N}, manje uzothola iphutha elingu-400. Ukucabanga okuguquguqukayo yikona kuphela imodi “yokucabanga” esekelwayo, futhi izilinganiso zangaphakathi ze-Anthropic zibonisa ukuthi isebenza kahle kakhulu kunebhajethi endala enwetshiwe noma kunjalo.

Ukucabanga okuguquguqukayo kuvaliwe ngokuzenzakalelayo, ngakho-ke izicelo ezingenayo insimu yokucabanga zisebenza ngaphandle komzila ocacile wokucabanga kwangaphakathi. Uma uhlelo lwakho lokusebenza luzuza emigqeni ecebile yemicabango—isibonelo, ukuhlela okuyinkimbinkimbi noma imisebenzi yokufaka amakhodi ezinyathelo eziningi—kufanele usethe ngokucacile ukucabanga: {type: “adaptive”} ukuze kuvuleke.

Esinye isengezo esikhulu yizabelomali zemisebenzi, okwamanje ezitholakala emphakathini ku-Claude Platform. Isabelomali somsebenzi sinikeza imodeli umgomo ojwayelekile wokuthi zingaki amathokheni okufanele asetshenziswe kulo lonke uhlelo lwe-ejensi: ukucabanga kwayo kwangaphakathi, izingcingo zamathuluzi, imiphumela yamathuluzi kanye nempendulo yokugcina. Imodeli ibona ukubala phansi okusebenzayo futhi kufanele ibeke phambili umsebenzi futhi iwuqede kahle njengoba isondela kwisabelomali, esikhundleni sokusebenza kuze kube yilapho ama-max_token eyinqamula ngokuzumayo.

Izabelomali zemisebenzi ziyiseluleko, hhayi imikhawulo eqinile, futhi zihlukile ngokomqondo kuma-max_token. I-max_tokens iyisilinganiso esiqinile ngesicelo ngasinye kumathokheni akhiqizwe futhi ayibonakali kumodeli, kuyilapho i-task_budget iyisilinganiso esithambile, esiqashelwa yimodeli kuyo yonke i-loop. Sebenzisa i-task_budget uma ufuna imodeli izilinganisele isifiso sayo ngokusekelwe kusibonelelo, futhi ugcine ama-max_tokens njengesivikelo sokuphepha ukuvimbela izindleko ezibalekayo.

Kuzodingeka uzame isabelomali somsebenzi ngomthwalo womsebenzi ngamunye: sibeke phansi kakhulu futhi imodeli ingase ithole isaphulelo kusenesikhathi noma iveze imiphumela engajulile; sibeke phezulu futhi uzokhokha amathokheni amaningi. I-Anthropic isikisela ukuthi ungasebenzisi isabelomali semisebenzi emisebenzini evulekile lapho ikhwalithi ibaluleke kakhulu; esikhundleni salokho, zigcinele imisebenzi lapho udinga khona ngempela imingcele ephezulu yokusetshenziswa kwezinsiza.

Izinguquko ze-API kanye ne-tokenizer ezithinta ukufuduka kusuka ku-Opus 4.6

Ngisho noma i-Opus 4.7 iyindlela yokuthuthukisa eqondile evela ku-Opus 4.6, kunezinguquko eziningana zezinga le-API kanye nomehluko we-tokenization okufanele uhlele kuwo ngokuphelele. Ukungazinaki lezi kungaholela ekudidekeni kwamaphutha angu-400 noma ekukhuphukeni kwezindleko okungalindelekile.

Okokuqala, i-Opus 4.7 isebenzisa i-tokenizer entsha efaka isandla ekusebenzeni kwayo okuthuthukisiwe kuyo yonke imisebenzi kodwa futhi ishintsha inani lamathokheni asetshenziswa yizinto ozifakile kanye nemiphumela yakho. Empeleni, umbhalo ofanayo ungasebenzisa cishe amathokheni angu-1.0-1.35× amaningi njengoba wenza ku-Opus 4.6—kuze kufike ku-35% ngaphezulu kuye ngohlobo lokuqukethwe. Ngakho-ke, i-/v1/messages/count_token endpoint izobika izinombolo ezihlukile ze-Opus 4.7 kunezamamodeli amadala.

Ukusebenza kahle kwethokheni kuzohluka ngesimo somsebenzi, kodwa izilinganiso zangaphakathi zokubhala ikhodi ze-Anthropic zisikisela ukuthi ukusetshenziswa kwethokheni ye-net kungathuthuka uma ucabangela ngokuhlakaniphile nokuhlela okufushane. Nokho, batusa ngokusobala ukuthi balinganise ithrafikhi yokukhiqiza yangempela futhi babuyekeze amapharamitha akho e-max_tokens ukuze banikeze i-headroom eyengeziwe, okuhlanganisa nanoma yiziphi iziqhubi zokucindezela ezenzakalelayo ongase uzilungiselele.

Okwesibili, i-Anthropic iqinise ukulawula kwamapharamitha okuthatha amasampula: ukuqala nge-Opus 4.7, ukusetha izinga lokushisa, i-top_p noma i-top_k kunoma yimaphi amanani angewona azenzakalelayo kuzobangela iphutha elingu-400. Indlela yokufuduka enconywayo iwukumane ulahle lawo mapharamitha ezicelweni zakho bese uncika ekuqondisweni kokuqondisa isitayela kanye nokunquma kwemodeli. Futhi noma ngabe usebenzise izinga lokushisa = 0 ngaphambilini ukuzama "ukumisa" izimpendulo, qaphela ukuthi ukunquma kweqiniso akukaze kuqinisekiswe.

Okwesithathu, okuqukethwe kokucabanga kukhishwe ezimpendulweni ngokuzenzakalelayo ku-Opus 4.7. Uzoqhubeka nokuthola amabhlogo okucabanga ezimpendulweni zokusakaza, kodwa insimu yabo yokucabanga izoba yize ngaphandle kokuthi ukhethe ngokusobala. Lolu shintsho oluthule luthuthukisa kancane ukubambezeleka futhi lunciphisa umkhawulokudonsa. Uma uhlelo lwakho lokusebenza ludinga izifinyezo ezifundekayo zokucabanga kwangaphakathi kwemodeli, ungakhetha ukungena futhi ngokusetha isibonisi ku-"summarized" ukuze kuphume ukucabanga ngokulungiswa kokucushwa komugqa owodwa.

Okokugcina, ezinye izinguquko zokuziphatha ezingezona izikhawu ze-API ezinzima zingadinga izibuyekezo ezisheshayo. Lokhu kufaka phakathi ukulandela imiyalelo engokoqobo (ikakhulukazi emazingeni aphansi omzamo), ama-subagent ambalwa azenzakalelayo, izingcingo ezimbalwa zamathuluzi ezizenzakalelayo, kanye noshintsho ekubuyekezweni kwenqubekela phambili ngesikhathi sokulandelwa kwesikhathi eside kwe-agent. I-Anthropic inikeza i- umhlahlandlela wokufuduka ngisho nomsizi wokufuduka ozenzakalelayo ngekhono le-Claude API lama-codebases asebenzisa i-Claude Code noma i-Agent SDK.

Ikhodi yeClaude ene-Opus 4.7: ukuthuthukiswa okusebenzayo komsebenzi wakho wokuthuthukisa

I-Claude Code, indawo yokubhala amakhodi ye-Anthropic, ihlolwe kakhulu nge-Opus 4.7, kanti uBoris Cherny—omunye wabadali bayo—ubelane ngezeluleko ezisekelwe emavikini okuyisebenzisa kumaphrojekthi angempela. Inguqulo emfushane: uma uzimisele ukushintsha indlela osebenza ngayo, ungathola intuthuko ebonakalayo ekukhiqizeni emisebenzini yobunjiniyela engathi sína.

Okokuqala, imodi entsha ezenzakalelayo isusa kakhulu ama-pop-up emvume ahlala njalo ayevame ukuphazamisa imisebenzi esebenze isikhathi eside. Esikhundleni sokukucela ukuthi uqinisekise konke ukuhlela noma umyalo wefayela, uClaude uhambisa lokho kuhlola kwemvume nge-classifier evumela ngokuzenzakalelayo izenzo eziphephile. Ungavumela i-Opus 4.7 ukuthi idlule ku-refactor, isebenzise izivivinyo noma ihlanze amafayela ngenkathi ugxila kokunye, bese ubuya uzobona ukuthi kufezile ini.

Imodi ezenzakalelayo okwamanje iyatholakala kubasebenzisi be-Max, Teams kanye ne-Enterprise futhi ingashintshwa ngokushesha nge-Shift+Tab emgqeni womyalo, noma ngokudonsela phansi kuhlelo lokusebenza lwedeskithophu kanye nesandiso se-VS Code. Inzuzo eyodwa engabhekwa njengebalulekile yikhono lokusebenzisa ama-ejenti amaningi ngesikhathi esisodwa, ngalinye lenza umsebenzi walo ojulile, bese ligxuma phakathi kwawo njengoba eqhubeka—ngaphandle kokunakekela yonke ingxoxo yokuvuma.

Uma ungakhululekile ukwenza okuzenzakalelayo ngokuphelele, iClaude Code inikeza ikhono le-/fewer-permission-prompts elihlaziya umlando wakho weseshini ukuthola imiyalo ephindaphindwayo nephephile esabangela izicelo zemvume. Ngemva kokuhlaziywa, kusikisela ukuthi imiyalo ongayifaka ohlwini oluyimfihlo ngokuphephile, unciphise ukuphazamiseka okuningi okuphansi ngenkathi ugcina izinyathelo ezinobungozi obukhulu ngemuva kokuvunyelwa.

Izifinyezo zingenye isici esihambisana ngokwemvelo nokugxila kwe-Opus 4.7 okuqondile. Uma ubuyela esimisweni esibe siqhubeka isikhathi eside, uClaude unikeza isifinyezo esifushane salokho akwenzile nalokho okusele. Lokhu kuyasiza kakhulu uma usuka emsebenzini oqhubekayo wokuhlaziya noma wokucwaninga; esikhundleni sokuskrola odongeni lwezingodo noma ama-diff, uthola isifinyezo esisheshayo sokuthi “nakhu lapho izinto zimi khona”.

Kulabo asebevele bethemba imodeli emisebenzini eyinkimbinkimbi, imodi yokugxila ifihla ingxoxo ephakathi futhi ibonise imiphumela yokugcina kuphela. UCherny ukhuluma ngokuyisebenzisa lapho engadingi ukubuka isinyathelo ngasinye u-Opus asithathayo—ufuna nje ikhodi eqediwe, izivivinyo noma amadokhumenti. Ungashintsha ukugxila ngomyalo we-/focus ngqo ku-CLI, okusiza ukunciphisa umsindo wokuqonda ngesikhathi seseshini yomsebenzi ojulile.

Ukulawula umzamo nakho kubalulekile ku-Claude Code: I-Opus 4.7 ilahla isabelomali sokucabanga esidala futhi ithembele ngokuphelele emzamweni wokuzivumelanisa nezimo. Ungalungisa umzamo nge-/effort, futhi wonke amazinga ngaphandle kwe-max ayaqhubeka phakathi nezikhathi. Iphethini yomuntu siqu kaCherny ukusebenzisa umzamo omkhulu kakhulu isikhathi esiningi, egcina inani eliphezulu kakhulu lezinkinga ezincane lapho yonke into eyengeziwe yokucabanga ibalulekile khona.

Kungashiwo ukuthi iseluleko esibaluleke kakhulu esivela kuCherny ukunikeza uClaude indlela yokuhlola umsebenzi wakhe. Kungakhathaliseki ukuthi i-stack ingakanani, i-ejenti idinga indlela yokusebenzisa izivivinyo noma ukugeleza kokuphela. Kumaphrojekthi angemuva, lokho kungasho iskripthi noma umyalo wokuzungeza iseva bese usebenzisa izivivinyo zokuhlanganisa; emsebenzini wangaphambili, uvame ukusebenzisa isandiso seChromium ukuze uClaude akwazi ukulawula isiphequluli, kuyilapho ikhono lokusebenzisa ikhompyutha limboza izinhlelo zokusebenza zedeskithophu.

Umsebenzi kaCherny uqobo uhlanganisa le filosofi ibe yikhono elenziwe ngokwezifiso lenza uClaude asebenzise izivivinyo, enze ikhodi ibe lula nge-/simplify, bese evula isicelo sokudonsa. Ngokwalokho akubona, lolu hlobo lwepayipi lokuqinisekisa bese uthumela luphindeka kabili noma kathathu kalula inani olithola ku-Opus 4.7. Luhambisana kahle nekhono elithuthukisiwe lemodeli lokuqinisekisa imiphumela yalo ngaphambi kokubika.

Kungani "ukusebenza kanzima kuqala, intshisekelo kamuva" isasebenza lapho kusetshenziswa amathuluzi amasha

Ngokuthakazelisayo, indlela abasebenzisi be-Anthropic kanye nabanamandla abakhuluma ngayo ngokuthola inzuzo evela ku-Opus 4.7 iphinda iphuzu elibanzi ngemisebenzi kanye nobuchwepheshe: intshisekelo ivame ukulandela ikhono, hhayi ukuliqala. UJensen Huang we-NVIDIA udume ngokugxeka isisho esithi “landela isifiso sakho,” ethi abantu abanikeza leso seluleko bavame ukuhlala ngokunethezeka, isikhathi eside ngemva kokuba sebedlule eminyakeni yokuqala enzima.

Isimo sengqondo esifanayo sivela lapho abantu bechaza indlela yokusebenza nomodeli ofana ne-Opus 4.7: khetha umzila lapho ungaba namandla khona ngempela, bese ujula isikhathi eside. Esikhundleni sokuphishekela i-hype, khetha isizinda—izinhlelo ze-backend, ukuzenzekela kwezokuphepha, ubunjiniyela bedatha, ama-devtools—lapho kunesidingo sangempela khona nalapho lo modeli ungathuthukisa khona amakhono akho eminyakeni eminingi, hhayi amasonto.

Leyo ncwadi yokudlala ibukeka ingathandeki uma uyibheka: ulinganisa intuthuko esikhundleni semizwa, wamukele ukuthi kuzoba nokungezwani, isithukuthezi kanye neziphazamiso, bese uqhubeka nokwenza ama-reps. Izinsuku ezinzima azisibo ubufakazi bokuthi ukhethe indlela engalungile; ngokuvamile ziwubufakazi bokuthi usebenza entweni ebalulekile. Ngokomongo we-Opus 4.7, lokho kusho ukuthi uhlaziya izixwayiso, amapayipi kanye nokugeleza kokubuyekezwa kuze kube yilapho ziqhubeka nokukhiqizwa, esikhundleni sokuphonsa imodeli kumaphrojekthi aseceleni angahleliwe bese ulindela umlingo.

Njengoba wakha ubuciko—bobabili ubuciko bakho kanye namathuluzi afana ne-Opus 4.7 abusekelayo—kancane kancane “uthola ilungelo lokuhlela” impilo yakho kanye nokubeka izinto ndawonye. Uthola ithuba lokukhetha izinkinga ezithakazelisayo, amaqembu angcono kanye nemingcele enempilo ngoba amakhono akho manje awatholakali kalula futhi athembekile. Ekuqaleni, ungase udele ibhalansi ukuze utshale imali eningi ekufundeni; kamuva, lokho kutshalwa kwezimali kukuvumela ukuthi uthenge isikhathi, ukulawula kanye nokuguquguquka.

Isifundo esiwusizo sabathuthukisi abasebenzisa iClaude Opus 4.7 silula: ungagcini nje ngokuhlela, kodwa zinikele. Phatha umsebenzi wokufuduka, ukulungiswa okusheshayo, umzamo kanye nokuhlolwa kwesabelomali semisebenzi njengeminyaka yokuzijwayeza ngemuva kwanoma yiliphi ikhono elibalulekile. Ngokuhamba kwesikhathi, yileyo ndlela ogcina ngayo unemisebenzi kanye nama-ejenti ashintsha ngempela indlela owakha ngayo isofthiwe, kunokuba ube nelinye ithuluzi "odlale ngalo impelasonto yonke."

Uma ihlanganiswa, iClaude Opus 4.7 iletha injini ebukhali, engokoqobo futhi enekhono kakhulu lokubhala amakhodi, umsebenzi wolwazi kanye nemisebenzi enzima yombono, ehlanganiswe nezilawuli ezingcono zokucabanga kanye nezindleko—futhi izungezwe yi-ecosystem efana neClaude Code eqondiswe emsebenzini wobunjiniyela wangempela, ongcolile. Abathuthukisi abathembela kulezi zinguquko, bacabange kabusha ngezeluleko zabo kanye nezindlela abazisebenzisayo, futhi banikeze imodeli izindlela eziqondile zokuqinisekisa umphumela wayo yibo okungenzeka kakhulu ukuthi bazizwe sengathi baqashe umlingani osebenza naye ongapheli, othanda imininingwane eminingi kunokuba nje athuthukiselwe ku-autocomplete ehlakaniphile kancane.

I-athikili ehlobene:

Ukusabalala kwe-API: izimbangela, izingozi kanye nendlela yokuthola kabusha ukulawula