{"id":659,"date":"2026-10-29T12:36:26","date_gmt":"2026-10-29T09:36:26","guid":{"rendered":"https:\/\/teknolata.com\/?p=659"},"modified":"2026-10-29T12:36:26","modified_gmt":"2026-10-29T09:36:26","slug":"google-gemini-4-carbonu-tanitti","status":"publish","type":"post","link":"https:\/\/teknolata.com\/index.php\/2026\/10\/29\/google-gemini-4-carbonu-tanitti\/","title":{"rendered":"Google Gemini 4 Carbon\u2019u Tan\u0131tt\u0131"},"content":{"rendered":"

Google, yeni nesil yapay zek\u00e2 modeli Gemini 4 Carbon\u2019u resmen duyurdu. Kodlama, karma\u015f\u0131k ak\u0131l y\u00fcr\u00fctme ve siber g\u00fcvenlik alanlar\u0131nda \u00f6nemli geli\u015ftirmeler i\u00e7eren model, Artificial Analysis Intelligence Index s\u0131ralamas\u0131nda zirveye yerle\u015fti. <\/p>\n

Yaz\u0131l\u0131mlardaki g\u00fcvenlik a\u00e7\u0131klar\u0131n\u0131 tespit edebilen, bunlar\u0131n ger\u00e7ekten istismar edilip edilemeyece\u011fini s\u0131nayan ve gerekli kod d\u00fczeltmelerini haz\u0131rlayan Gemini 4 Carbon, ilk etapta s\u0131n\u0131rl\u0131 say\u0131daki siber g\u00fcvenlik uzman\u0131n\u0131n kullan\u0131m\u0131na a\u00e7\u0131l\u0131yor.<\/p>\n

Google\u2019\u0131n iki hafta \u00f6nce kullan\u0131ma sundu\u011fu Gemini 4 Argon modelinin ard\u0131ndan gelen Gemini 4 Carbon, \u00f6zellikle uzun s\u00fcreli ve \u00e7ok a\u015famal\u0131 g\u00f6revlerde daha y\u00fcksek performans sunuyor. Artificial Analysis Intelligence Index\u2019in g\u00fcncellenen 5.0 s\u00fcr\u00fcm\u00fcnde 59 puana ula\u015fan model, genel yapay zek\u00e2 performans s\u0131ralamas\u0131nda birinci s\u0131raya yerle\u015fti. Kodlama, otonom g\u00f6rev y\u00fcr\u00fctme, bilimsel \u00e7al\u0131\u015fmalar ve \u00e7ok modlu i\u00e7eriklerin anla\u015f\u0131lmas\u0131 gibi alanlarda da geli\u015ftirmeler yap\u0131ld\u0131.<\/p>\n

\n

<\/figure>\n<\/p><\/div>\n

Yeni modelin \u00f6nemli \u00f6zelliklerinden biri, karma\u015f\u0131k m\u00fchendislik problemleri \u00fczerinde g\u00fcnlerce \u00e7al\u0131\u015fabilmesi. Gemini 4 Carbon, b\u00fcy\u00fck yaz\u0131l\u0131m projelerindeki hatalar\u0131n temel nedenlerini ara\u015ft\u0131rabiliyor, farkl\u0131 \u00e7\u00f6z\u00fcm yollar\u0131n\u0131 deneyebiliyor ve ba\u015far\u0131s\u0131z giri\u015fimlerin ard\u0131ndan \u00e7al\u0131\u015fmaya devam edebiliyor. <\/p>\n

Google\u2019\u0131n \u015firket i\u00e7i testlerinde modelin, b\u00fcy\u00fck kod tabanlar\u0131 ve d\u00fc\u015f\u00fck seviyeli sistemlerdeki karma\u015f\u0131k hatalar\u0131 bulmak ve d\u00fczeltmek i\u00e7in g\u00fcnler s\u00fcren otonom i\u015f ak\u0131\u015flar\u0131n\u0131 tamamlad\u0131\u011f\u0131 g\u00f6r\u00fcld\u00fc.<\/p>\n


<\/figure>\n

Yaz\u0131l\u0131m geli\u015ftirme taraf\u0131nda tam kapsaml\u0131 web uygulamalar\u0131 olu\u015fturma ve \u00fcretilen \u00e7\u0131kt\u0131lar\u0131n kalitesinde de ilerleme kaydedildi. Gemini 4 Carbon, WebDev Arena testinde \u00f6nceki performans\u0131n\u0131 89 Elo puan\u0131 art\u0131rarak 1768 puana ula\u015ft\u0131 ve d\u00f6rd\u00fcnc\u00fc s\u0131raya y\u00fckseldi. Modelin bilimsel g\u00f6revlerde, farkl\u0131 ara\u00e7lar\u0131n birlikte kullan\u0131ld\u0131\u011f\u0131 i\u015flemlerde ve g\u00f6rsel ile metin gibi birden fazla veri t\u00fcr\u00fcn\u00fc i\u00e7eren \u00e7al\u0131\u015fmalarda da performans\u0131 geli\u015ftirildi.<\/p>\n

Gemini 4 Carbon\u2019un en dikkat \u00e7ekici sonu\u00e7lar\u0131 ise siber g\u00fcvenlik testlerinde ortaya \u00e7\u0131kt\u0131. Model, ger\u00e7ek yaz\u0131l\u0131mlardaki g\u00fcvenlik a\u00e7\u0131klar\u0131n\u0131 tespit etme becerisinin \u00f6l\u00e7\u00fcld\u00fc\u011f\u00fc SEC-bench Pro testinde y\u00fczde 88,2 ba\u015far\u0131 oran\u0131na ula\u015ft\u0131. Bu test, yaz\u0131l\u0131m projelerinin i\u00e7erisinde bulunan karma\u015f\u0131k g\u00fcvenlik a\u00e7\u0131klar\u0131n\u0131n otonom \u015fekilde belirlenmesini de\u011ferlendiriyor.<\/p>\n


<\/figure>\n

Tespit edilen a\u00e7\u0131klar\u0131n ger\u00e7ekten istismar edilip edilemeyece\u011fini s\u0131nayan ExploitGym testinde y\u00fczde 54,8 ba\u015far\u0131 elde edildi. G\u00fcvenlik a\u00e7\u0131klar\u0131n\u0131n temel nedenlerini ortadan kald\u0131ran kod d\u00fczeltmelerinin de\u011ferlendirildi\u011fi Collinear CWE-bench testinde ise y\u00fczde 71 ba\u015far\u0131 oran\u0131na ula\u015f\u0131ld\u0131. <\/p>\n

\u00dc\u00e7 testte de yeni model, duyuru s\u0131ras\u0131nda alan\u0131ndaki en y\u00fcksek sonu\u00e7lar\u0131 elde etti. S\u00f6z konusu harici test sonu\u00e7lar\u0131, modelin g\u00fcvenlik k\u0131s\u0131tlamalar\u0131 uygulanmadan ger\u00e7ekle\u015ftirilen de\u011ferlendirmelerine dayan\u0131yor.<\/p>\n


<\/figure>\n

Google\u2019\u0131n \u015firket i\u00e7i g\u00fcvenli kod \u00fcretme testlerinde Gemini 4 Carbon, Gemini 4 Argon\u2019un ba\u015far\u0131 oran\u0131n\u0131 iki kattan fazla art\u0131rd\u0131. Chrome ekibinin ger\u00e7ekle\u015ftirdi\u011fi de\u011ferlendirmelerde ise model, y\u00fcksek kaliteli g\u00fcvenlik yamalar\u0131 \u00fcretirken ba\u015fka bir ara\u015ft\u0131rma laboratuvar\u0131n\u0131n \u00fcst d\u00fczey yapay zek\u00e2 modelinden 10 kat\u0131n \u00fczerinde daha iyi sonu\u00e7 verdi. <\/p>\n

Google Cloud b\u00fcnyesindeki siber g\u00fcvenlik platformu Wiz de web uygulamalar\u0131na y\u00f6nelik s\u0131zma testi de\u011ferlendirmelerinde Carbon\u2019un liderli\u011fini kaydetti. Bu yetenekler, Wiz\u2019in Scan for Good giri\u015fiminde kullan\u0131lan g\u00fcvenlik testleri a\u00e7\u0131s\u0131ndan da \u00f6nem ta\u015f\u0131yor.<\/p>\n


<\/figure>\n

G\u00fcvenlik a\u00e7\u0131klar\u0131n\u0131n belirlenmesi, do\u011frulanmas\u0131 ve d\u00fczeltilmesinin yan\u0131 s\u0131ra modelin kendi \u00e7al\u0131\u015fma g\u00fcvenli\u011fi \u00fczerinde de de\u011fi\u015fiklikler yap\u0131ld\u0131. Gemini 4 Argon\u2019da kullan\u0131lan \u00e7ok katmanl\u0131 g\u00fcvenlik altyap\u0131s\u0131 Carbon\u2019da da korunuyor. <\/p>\n

Bu altyap\u0131, model e\u011fitimi s\u0131ras\u0131nda uygulanan g\u00fcvenlik \u00f6nlemlerini, ger\u00e7ek zamanl\u0131 k\u00f6t\u00fcye kullan\u0131m engellemelerini, prompt injection sald\u0131r\u0131lar\u0131na kar\u015f\u0131 korumalar\u0131 ve yapay zek\u00e2n\u0131n ak\u0131l y\u00fcr\u00fctme s\u00fcre\u00e7lerinin izlenmesini kaps\u0131yor.<\/p>\n

Modelin kendisine verilen g\u00f6revleri yerine getirirken kestirme yollara ba\u015fvurmas\u0131n\u0131 ve de\u011ferlendirme sistemlerindeki a\u00e7\u0131klardan yararlanmas\u0131n\u0131 engellemek i\u00e7in de ek \u00e7al\u0131\u015fmalar ger\u00e7ekle\u015ftirildi. Terminal-Bench 4.0 testinde ba\u015far\u0131l\u0131 hile giri\u015fimlerinin oran\u0131 y\u00fczde 0\u2019a d\u00fc\u015f\u00fcr\u00fcld\u00fc. <\/p>\n

Hile yapma giri\u015fimlerinin toplam oran\u0131 ise y\u00fczde 0,4\u2019e geriledi. Bu sonu\u00e7lar elde edilirken modelin genel do\u011fruluk performans\u0131 da art\u0131r\u0131ld\u0131. Google\u2019\u0131n ger\u00e7ek kodlama ortamlar\u0131nda ger\u00e7ekle\u015ftirdi\u011fi \u015firket i\u00e7i testlerde y\u00fcksek \u00f6nem derecesine sahip uyumsuz davran\u0131\u015f uyar\u0131lar\u0131 yakla\u015f\u0131k y\u00fczde 80 azald\u0131.<\/p>\n

Gemini 4 Carbon\u2019da yapay zek\u00e2 modellerinin kar\u015f\u0131la\u015ft\u0131\u011f\u0131 \u00f6nemli sorunlardan biri olan ger\u00e7e\u011fe dayanmayan bilgi \u00fcretimi konusunda da iyile\u015ftirmeler bulunuyor. \u015eirket i\u00e7i de\u011ferlendirmelerde modelin yeterli kan\u0131t olmadan ileri s\u00fcrd\u00fc\u011f\u00fc iddialar\u0131n say\u0131s\u0131, Gemini 4 Argon\u2019a k\u0131yasla yakla\u015f\u0131k yar\u0131ya indirildi. <\/p>\n

Model, do\u011frulayamad\u0131\u011f\u0131 bilgileri kesin ifadelerle sunmak yerine kullan\u0131lan y\u00f6ntemin s\u0131n\u0131rlamalar\u0131n\u0131 belirtebiliyor. Bir i\u015flemi tamamlamak i\u00e7in gerekli ko\u015fullar sa\u011flanmad\u0131\u011f\u0131nda ise ger\u00e7ekle\u015fmemi\u015f bir i\u015flemi tamamlanm\u0131\u015f gibi g\u00f6stermeden hangi ad\u0131mlar\u0131n eksik kald\u0131\u011f\u0131n\u0131 a\u00e7\u0131klayabiliyor.<\/p>\n

Kullan\u0131c\u0131larla birlikte \u00e7al\u0131\u015fma konusunda da de\u011fi\u015fiklikler yap\u0131ld\u0131. Gemini 4 Carbon, belirsiz talimatlarla kar\u015f\u0131la\u015ft\u0131\u011f\u0131nda do\u011frudan varsay\u0131mlarda bulunmak yerine ilgili kaynaklar\u0131 ara\u015ft\u0131rabiliyor ve elde etti\u011fi bilgiler aras\u0131ndaki tutars\u0131zl\u0131klar\u0131 kullan\u0131c\u0131ya iletebiliyor. Uzun ve \u00e7ok a\u015famal\u0131 g\u00f6revlerde belirlenen s\u0131n\u0131rlar\u0131n d\u0131\u015f\u0131na \u00e7\u0131kma e\u011filimi y\u00fczde 26 azalt\u0131l\u0131rken, kullan\u0131c\u0131 taraf\u0131ndan istenen i\u015flemleri eksik yerine getirme oran\u0131 y\u00fczde 29 d\u00fc\u015f\u00fcr\u00fcld\u00fc.<\/p>\n

Prompt injection sald\u0131r\u0131lar\u0131na kar\u015f\u0131 dayan\u0131kl\u0131l\u0131k da geli\u015ftirildi. Bu sald\u0131r\u0131larda k\u00f6t\u00fc niyetli ki\u015filer, yapay zek\u00e2n\u0131n i\u015fledi\u011fi belgeler veya d\u0131\u015f kaynaklar i\u00e7erisine zararl\u0131 talimatlar yerle\u015ftirerek modelin davran\u0131\u015f\u0131n\u0131 de\u011fi\u015ftirmeye \u00e7al\u0131\u015f\u0131yor. <\/p>\n

Gemini 4 Argon, Gray Swan Indirect Prompt Injection Benchmark testinde y\u00fczde 0,7 sald\u0131r\u0131 ba\u015far\u0131 oran\u0131na sahipken Gemini 4 Carbon\u2019da bu oran y\u00fczde 0\u2019a geriledi. Carbon, de\u011ferlendirmede kullan\u0131lan sald\u0131r\u0131 giri\u015fimlerinin tamam\u0131na kar\u015f\u0131 koydu.<\/p>\n

Gemini 4 Carbon\u2019un da\u011f\u0131t\u0131m\u0131 ise \u00f6nceki Gemini modellerinden farkl\u0131 \u015fekilde ilerliyor. Google, yeni modeli do\u011frudan b\u00fct\u00fcn kullan\u0131c\u0131lar\u0131n eri\u015fimine a\u00e7mak yerine a\u015famal\u0131 bir da\u011f\u0131t\u0131m s\u00fcreci uyguluyor. <\/p>\n

\u015eirket, modelin genel kullan\u0131ma sunulmas\u0131ndan \u00f6nce ABD h\u00fck\u00fcmetinin g\u00f6n\u00fcll\u00fc model test program\u0131na kat\u0131l\u0131yor. Bu s\u00fcre\u00e7te modelin g\u00fcvenlik kapasitesi ve olas\u0131 k\u00f6t\u00fcye kullan\u0131m riskleri de\u011ferlendiriliyor.<\/p>\n

\u0130lk eri\u015fim hakk\u0131, 28 Ekim haftas\u0131ndan itibaren se\u00e7ilmi\u015f siber savunma uzmanlar\u0131na veriliyor. Ard\u0131ndan Google\u2019\u0131n Fairwind Program\u0131 kapsam\u0131nda daha fazla g\u00fcvenilir i\u015f orta\u011f\u0131na eri\u015fim sa\u011flanacak. <\/p>\n

\u0130lk kullan\u0131c\u0131lardan gelen geri bildirimlerle g\u00fcvenlik mekanizmalar\u0131 \u00fczerinde d\u00fczenlemeler yap\u0131lacak. Geli\u015ftiriciler, i\u015fletmeler ve bireysel kullan\u0131c\u0131lar i\u00e7in genel eri\u015fimin ne zaman ba\u015flayaca\u011f\u0131na ili\u015fkin kesin bir tarih hen\u00fcz a\u00e7\u0131klanmad\u0131.<\/p>\n

Gemini 4 Carbon\u2019un fiyatland\u0131rmas\u0131 da belli oldu. Model eri\u015fime a\u00e7\u0131ld\u0131\u011f\u0131nda ba\u015flang\u0131\u00e7 d\u00f6neminde her 1 milyon giri\u015f token\u0131 i\u00e7in 2 dolar, her 1 milyon \u00e7\u0131k\u0131\u015f token\u0131 i\u00e7in ise 10 dolar \u00fccret al\u0131nacak. <\/p>\n

\u00d6nbelle\u011fe al\u0131nan giri\u015f tokenlar\u0131nda standart giri\u015f fiyat\u0131 \u00fczerinden y\u00fczde 95 indirim uygulanacak. B\u00f6ylece \u00f6nbellekten kullan\u0131lan 1 milyon giri\u015f token\u0131n\u0131n maliyeti 0,10 dolar olacak.<\/p>\n

Ba\u015flang\u0131\u00e7 fiyatland\u0131rmas\u0131n\u0131n sona ermesinin ard\u0131ndan her 1 milyon giri\u015f token\u0131n\u0131n fiyat\u0131 4 dolara, her 1 milyon \u00e7\u0131k\u0131\u015f token\u0131n\u0131n fiyat\u0131 ise 20 dolara y\u00fckselecek. B\u00f6ylece yeni model, ba\u015flang\u0131\u00e7 d\u00f6neminde Gemini 4 Argon ile ayn\u0131 fiyat \u00fczerinden sunulacak. \u0130ndirimli fiyatland\u0131rman\u0131n ne kadar s\u00fcre ge\u00e7erli olaca\u011f\u0131na ili\u015fkin bir tarih payla\u015f\u0131lmad\u0131.<\/p>\n

TAGS: <\/strong>Ger\u00e7ek<\/a><\/span>Giri\u015f<\/a><\/span>G\u00fcvenlik<\/a><\/span>Modeli<\/a><\/span>Test<\/a> <\/div>\n","protected":false},"excerpt":{"rendered":"

Google Gemini 4 Carbon’u tan\u0131tt\u0131. Siber g\u00fcvenlikte dikkat \u00e7eken test sonu\u00e7lar\u0131 payla\u015f\u0131l\u0131rken modelin eri\u015fim takvimi ve API fiyatlar\u0131 da a\u00e7\u0131kland\u0131.<\/p>\n","protected":false},"author":1,"featured_media":660,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[794,795,792,791,793],"class_list":["post-659","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tumu","tag-gercek","tag-giris","tag-guvenlik","tag-modeli","tag-test"],"_links":{"self":[{"href":"https:\/\/teknolata.com\/index.php\/wp-json\/wp\/v2\/posts\/659","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/teknolata.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/teknolata.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/teknolata.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/teknolata.com\/index.php\/wp-json\/wp\/v2\/comments?post=659"}],"version-history":[{"count":1,"href":"https:\/\/teknolata.com\/index.php\/wp-json\/wp\/v2\/posts\/659\/revisions"}],"predecessor-version":[{"id":666,"href":"https:\/\/teknolata.com\/index.php\/wp-json\/wp\/v2\/posts\/659\/revisions\/666"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/teknolata.com\/index.php\/wp-json\/wp\/v2\/media\/660"}],"wp:attachment":[{"href":"https:\/\/teknolata.com\/index.php\/wp-json\/wp\/v2\/media?parent=659"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/teknolata.com\/index.php\/wp-json\/wp\/v2\/categories?post=659"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/teknolata.com\/index.php\/wp-json\/wp\/v2\/tags?post=659"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}