{"id":174740,"date":"2026-07-23T19:16:20","date_gmt":"2026-07-23T17:16:20","guid":{"rendered":"https:\/\/jfrog.com\/blog\/why-model-routing-backfires\/"},"modified":"2026-09-17T06:59:06","modified_gmt":"2026-09-17T04:59:06","slug":"why-model-routing-backfires","status":"publish","type":"post","link":"https:\/\/jfrog.com\/fr\/blog\/why-model-routing-backfires\/","title":{"rendered":"Pourquoi le routage de mod\u00e8les s\u2019av\u00e8re contre-productif et comment cr\u00e9er des agents qui n\u2019\u00e9puisent pas votre budget"},"content":{"rendered":"<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-170400 size-full\" src=\"https:\/\/jfrog.com\/wp-content\/uploads\/2026\/07\/863x300-9-2.png\" alt=\"\" width=\"863\" height=\"300\" srcset=\"https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2026\/07\/863x300-9-2.png?speedsize=w_863 863w, https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2026\/07\/863x300-9-2.png?speedsize=w_300 300w, https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2026\/07\/863x300-9-2.png?speedsize=w_768 768w\" sizes=\"auto, (max-width: 863px) 100vw, 863px\" \/><\/p>\n<p>Le routage de mod\u00e8les promet de r\u00e9duire vos d\u00e9penses en agents d\u2019IA en d\u00e9chargeant les t\u00e2ches courantes sur des mod\u00e8les moins chers comme Claude&#160;Haiku, tout en r\u00e9servant les mod\u00e8les de pointe comme Claude&#160;Sonnet au raisonnement complexe. Dans la bonne configuration, les strat\u00e9gies de routage peuvent r\u00e9duire les co\u00fbts d\u2019inf\u00e9rence de <a href=\"https:\/\/proceedings.iclr.cc\/paper_files\/paper\/2025\/hash\/5503a7c69d48a2f86fc00b3dc09de686-Abstract-Conference.html\">40 \u00e0 85&#160;%<\/a>. Mais si vous impl\u00e9mentez de mani\u00e8re incorrecte le routage au sein d\u2019un agent multi-tour, vous risquez de payer plus cher que si vous n\u2019aviez jamais effectu\u00e9 de routage. Voici pourquoi et comment le corriger.<\/p>\n<h2>Comment les co\u00fbts des jetons s\u2019accumulent au cours d\u2019une session<\/h2>\n<p>Comprendre pourquoi le routage s\u2019av\u00e8re contre-productif n\u00e9cessite de comprendre comment les sessions agentiques accumulent des co\u00fbts. Chaque \u00e9change au sein d\u2019une conversation augmente de mani\u00e8re cumulative le nombre de jetons, et cet effet cumulatif est le moteur \u00e0 la fois de vos \u00e9conomies et de vos risques.<\/p>\n<p>Chaque \u00e9change dans une session d\u2019agent comporte une charge utile de jetons croissante. Le tableau ci-dessous montre comment la distribution \u00e9volue au cours d&rsquo;une session typique de trois tours&#160;:<\/p>\n<table>\n<tbody>\n<tr>\n<td><strong>\u00c9change<\/strong><\/td>\n<td><strong>Nouveaux jetons d&rsquo;entr\u00e9e<\/strong><\/td>\n<td><strong>Jetons mis en cache<\/strong><\/td>\n<\/tr>\n<tr>\n<td>Tour&#160;1<\/td>\n<td>Haut (message de base)<\/td>\n<td>Aucun<\/td>\n<\/tr>\n<tr>\n<td>Tour&#160;2<\/td>\n<td>Faible (message de suivi)<\/td>\n<td>Haut (tout le tour&#160;1)<\/td>\n<\/tr>\n<tr>\n<td>Tour&#160;3+<\/td>\n<td>Minimal<\/td>\n<td>Tr\u00e8s \u00e9lev\u00e9 (effet cumul\u00e9)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>&#160;<\/p>\n<p>D\u00e8s le tour&#160;3, les jetons mis en cache constituent la grande majorit\u00e9 de votre charge utile de jetons. C\u2019est l\u2019efficacit\u00e9 que vous cherchez \u00e0 prot\u00e9ger. Changer de mod\u00e8le en cours de session d\u00e9truit enti\u00e8rement cette protection.<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-170384 size-full\" src=\"https:\/\/jfrog.com\/wp-content\/uploads\/2026\/07\/image4.png\" alt=\"\" width=\"960\" height=\"540\" srcset=\"https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2026\/07\/image4.png?speedsize=w_960 960w, https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2026\/07\/image4.png?speedsize=w_300 300w, https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2026\/07\/image4.png?speedsize=w_768 768w\" sizes=\"auto, (max-width: 960px) 100vw, 960px\" \/><\/p>\n<h2>Pourquoi changer de mod\u00e8le en cours de session augmente-t-il vos co\u00fbts&#160;?<\/h2>\n<p>Changer de mod\u00e8le LLM en cours de session annule les \u00e9conomies de cache de prompts accumul\u00e9es, car le cache de chaque fournisseur est propre au mod\u00e8le. Le nouveau mod\u00e8le n\u2019a pas acc\u00e8s \u00e0 l\u2019historique stock\u00e9 du mod\u00e8le pr\u00e9c\u00e9dent et doit relire l\u2019int\u00e9gralit\u00e9 de la conversation depuis le d\u00e9but au tarif standard des jetons d\u2019entr\u00e9e.<\/p>\n<p>Les caches de prompts fonctionnent par correspondance de pr\u00e9fixes. Le fournisseur de LLM stocke un hachage du pr\u00e9fixe de votre message. Tant que le pr\u00e9fixe reste stable et que le mod\u00e8le reste le m\u00eame, chaque \u00e9change suivant b\u00e9n\u00e9ficie de tarifs fortement r\u00e9duits pour les jetons mis en cache. <a href=\"https:\/\/www.anthropic.com\/claude\/sonnet\" target=\"_blank\" rel=\"noopener\">Anthropic<\/a>, par exemple, offre jusqu\u2019\u00e0 90&#160;% de r\u00e9duction sur les jetons d\u2019entr\u00e9e mis en cache. D\u00e8s que vous changez de mod\u00e8le, le nouveau mod\u00e8le d\u00e9marre avec un cache froid. Il traite l\u2019ensemble de l\u2019historique de la conversation au co\u00fbt total des jetons d\u2019entr\u00e9e, comme si le tour&#160;1 n\u2019avait jamais eu lieu. Pour les sessions courtes, c&rsquo;est un d\u00e9sagr\u00e9ment mineur. Pour les workflows de codage agentique o\u00f9 les fen\u00eatres de contexte peuvent s\u2019\u00e9tendre jusqu\u2019\u00e0 50&#160;000&#160;jetons ou plus, un seul changement de mod\u00e8le en cours de session peut faire grimper les co\u00fbts au point d\u2019\u00e9liminer toutes les \u00e9conomies que le routage \u00e9tait cens\u00e9 apporter.<\/p>\n<h2>Ce que vous ne devez pas faire<\/h2>\n<p>Plusieurs sch\u00e9mas de routage courants ressemblent \u00e0 des optimisations, mais nuisent activement \u00e0 votre profil de co\u00fbts. Voici les causes les plus fr\u00e9quentes d\u2019augmentations inattendues de co\u00fbts dans les syst\u00e8mes d\u2019agents en production&#160;:<\/p>\n<ul>\n<li aria-level=\"1\"><strong>Changer de mod\u00e8le au milieu d\u2019une longue conversation effectuer de r\u00e9sum\u00e9&#160;:<\/strong> cela oblige le nouveau mod\u00e8le \u00e0 retraiter l\u2019ensemble de l\u2019historique de la conversation au co\u00fbt total des jetons d\u2019entr\u00e9e. Le co\u00fbt d\u2019un d\u00e9faut de cache pour une session de 20&#160;000&#160;jetons peut \u00eatre sup\u00e9rieur aux \u00e9conomies que le changement devait g\u00e9n\u00e9rer.<\/li>\n<li aria-level=\"1\"><strong>Activer le routage automatique en cours de session dans les passerelles LLM&#160;:<\/strong> les outils de passerelle LLM tels que LiteLLM, OpenRouter et Portkey prennent en charge le routage automatique en fonction de la complexit\u00e9 de la requ\u00eate. C\u2019est puissant pour les requ\u00eates \u00e0 un seul \u00e9change, mais l\u2019activer en milieu de session dans une boucle agentique invalide votre cache&#160;KV \u00e0 chaque limite de routage.<\/li>\n<li aria-level=\"1\"><strong>Modifier des outils ou des prompts syst\u00e8me en cours de session&#160;:<\/strong> les d\u00e9finitions d\u2019outils se trouvent dans le pr\u00e9fixe mis en cache. L&rsquo;ajout ou la suppression d&rsquo;un outil au cours d&rsquo;une session invalide l&rsquo;ensemble du cache en aval de ce changement. Consid\u00e9rez la configuration d\u2019outil de session comme immuable une fois d\u00e9marr\u00e9e.<\/li>\n<li aria-level=\"1\"><strong>Utiliser des appels de compression avec un mod\u00e8le distinct&#160;:<\/strong> de nombreux agents compressent l\u2019historique long en g\u00e9n\u00e9rant un appel de r\u00e9sum\u00e9 avec un mod\u00e8le peu co\u00fbteux. Si cet appel utilise un mod\u00e8le diff\u00e9rent ou un prompt syst\u00e8me diff\u00e9rent, il g\u00e9n\u00e8re son propre d\u00e9faut de cache et r\u00e9initialise le cache de la session principale au retour.<\/li>\n<\/ul>\n<p>\u00c9viter ces sch\u00e9mas n\u00e9cessite une conception de session intentionnelle, et pas seulement une logique de routage correcte. Votre strat\u00e9gie de routage ne vaut que ce que vaut votre architecture de session.<\/p>\n<h2>Que faire \u00e0 la place<\/h2>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-170385 size-full\" src=\"https:\/\/jfrog.com\/wp-content\/uploads\/2026\/07\/image2-1.png\" alt=\"\" width=\"960\" height=\"540\" srcset=\"https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2026\/07\/image2-1.png?speedsize=w_960 960w, https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2026\/07\/image2-1.png?speedsize=w_300 300w, https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2026\/07\/image2-1.png?speedsize=w_768 768w\" sizes=\"auto, (max-width: 960px) 100vw, 960px\" \/><\/p>\n<p>Le routage efficace des mod\u00e8les dans les syst\u00e8mes agentiques rel\u00e8ve du niveau de la session, et non du niveau de la requ\u00eate. Les pratiques suivantes vous permettent de b\u00e9n\u00e9ficier des avantages financiers du routage sans sacrifier les \u00e9conomies li\u00e9es au cache de prompts qui rendent l\u2019IA agentique \u00e9conomiquement viable&#160;:<\/p>\n<ol>\n<li aria-level=\"1\"><strong>Minimiser le nombre d\u2019\u00e9changes par t\u00e2che et d\u00e9l\u00e9guer t\u00f4t&#160;:<\/strong> maintenez les sessions d\u2019agent individuelles aussi courtes que possible. Lorsque vous savez qu\u2019une t\u00e2che n\u00e9cessite un contexte \u00e9tendu (par exemple, une refactorisation complexe de plusieurs fichiers, une v\u00e9rification de conformit\u00e9 de longue dur\u00e9e), d\u00e9l\u00e9guez-la \u00e0 un sous-agent d\u00e8s le d\u00e9but, avant que la session n\u2019accumule de l\u2019historique. Des sessions courtes signifient de petits caches, ce qui signifie qu\u2019un changement de mod\u00e8le co\u00fbte moins cher.<\/li>\n<li aria-level=\"1\"><strong>Utiliser le sch\u00e9ma de sous-agent pour la diversit\u00e9 des mod\u00e8les&#160;:<\/strong> au lieu de changer de mod\u00e8le au sein d\u2019une m\u00eame session, cr\u00e9ez un agent orchestrateur qui d\u00e9l\u00e8gue les sous-t\u00e2ches \u00e0 des sous-agents sp\u00e9cialis\u00e9s. Chaque sous-agent repart de z\u00e9ro avec uniquement le contexte dont il a besoin et s&rsquo;ex\u00e9cute sur le mod\u00e8le le mieux adapt\u00e9 \u00e0 sa t\u00e2che sp\u00e9cifique. Aucune invalidation du cache. Aucune p\u00e9nalit\u00e9 d\u2019historique accumul\u00e9. Ce sch\u00e9ma (parfois appel\u00e9 l\u2019architecture orchestrateur-sous-agent) est le moyen le plus propre d\u2019obtenir une diversit\u00e9 de mod\u00e8les bas\u00e9e sur les co\u00fbts sans subir la p\u00e9nalit\u00e9 d\u2019un changement en cours de session.<\/li>\n<li aria-level=\"1\"><strong>R\u00e9sumer d\u00e9lib\u00e9r\u00e9ment avant tout changement de mod\u00e8le requis&#160;:<\/strong> si un changement de mod\u00e8le en cours de session est in\u00e9vitable, cr\u00e9ez un passage de relais structur\u00e9&#160;: r\u00e9sumez explicitement le contexte actif, l\u2019\u00e9tat actuel de la t\u00e2che et les d\u00e9cisions en suspens dans un message compact. D\u00e9marrez une nouvelle session avec le nouveau mod\u00e8le en utilisant uniquement ce r\u00e9sum\u00e9 comme entr\u00e9e. Vous payez des co\u00fbts d\u2019entr\u00e9e standard sur un petit r\u00e9sum\u00e9 au lieu de co\u00fbts mis en cache sur une relecture d\u2019historique de 20&#160;000&#160;jetons.<\/li>\n<li aria-level=\"1\"><strong>Verrouiller la configuration de votre session&#160;:<\/strong> les prompts syst\u00e8me, les d\u00e9finitions d\u2019outils et la s\u00e9lection du mod\u00e8le doivent \u00eatre fix\u00e9s au d\u00e9but de la session. Consid\u00e9rez-les comme un sch\u00e9ma&#160;: les modifier en cours de session constitue un \u00e9v\u00e9nement de migration, et non un ajustement de configuration. Si votre framework d\u2019agents permet le chargement dynamique d\u2019outils, auditez o\u00f9 et quand ces chargements se produisent par rapport \u00e0 votre limite de pr\u00e9fixe de cache.<\/li>\n<\/ol>\n<h2>Comment JFrog&#160;Boost prot\u00e8ge automatiquement l\u2019efficacit\u00e9 de votre cache<\/h2>\n<p>Bien que la gestion de la logique de routage des mod\u00e8les soit essentielle \u00e0 la p\u00e9rennit\u00e9 des agents, la surcharge de contexte agit comme un v\u00e9ritable multiplicateur de force pour le gaspillage de jetons. Chaque fois qu\u2019un agent ex\u00e9cute un build, lance une suite de tests ou r\u00e9cup\u00e8re des logs, des centaines de lignes de bruit r\u00e9p\u00e9titif du terminal inondent l\u2019historique de la conversation.<\/p>\n<p>Dans une session multi-tour, ce bruit s\u2019accumule rapidement. Cela transforme ce qui aurait d\u00fb \u00eatre un historique l\u00e9ger en une charge utile de 50&#160;000&#160;jetons en quelques tours. Cela nuit directement \u00e0 votre ing\u00e9nierie des co\u00fbts&#160;: si votre session est encombr\u00e9e de bruits de terminal, la \u00ab&#160;p\u00e9nalit\u00e9 li\u00e9e aux d\u00e9fauts de cache&#160;\u00bb dont nous avons parl\u00e9 pr\u00e9c\u00e9demment devient exponentiellement plus co\u00fbteuse d\u00e8s qu\u2019un changement de mod\u00e8le intervient en milieu de session.<\/p>\n<p>Lorsqu\u2019ils utilisent des agents de codage, les d\u00e9veloppeurs ne devraient pas avoir \u00e0 filtrer manuellement la sortie du terminal simplement pour \u00e9viter qu\u2019une \u00e9tape de build bruyante ne gonfle leur facture de LLM.<\/p>\n<p><strong><a href=\"https:\/\/boost.jfrog.com\/\">JFrog&#160;Boost<\/a> est con\u00e7u pour r\u00e9soudre cette surcharge de contexte directement au sein du workflow de d\u00e9veloppement. <\/strong>Boost est un outil CLI l\u00e9ger qui compacte intelligemment la sortie du terminal lors des boucles agentiques, en \u00e9liminant les \u00e9l\u00e9ments r\u00e9p\u00e9titifs tout en pr\u00e9servant parfaitement le signal critique, comme les traces de pile d\u2019erreurs exactes. Au lieu d\u2019envoyer des milliers de jetons de logs de build bruts \u00e0 votre mod\u00e8le, Boost all\u00e8ge la charge utile avant m\u00eame qu\u2019elle n\u2019atteigne l\u2019agent.<\/p>\n<p>En conservant un historique de session all\u00e9g\u00e9, Boost r\u00e9duit consid\u00e9rablement le nombre de jetons de base trait\u00e9s \u00e0 chaque \u00e9change. Cela minimise la p\u00e9nalit\u00e9 financi\u00e8re des changements in\u00e9vitables de routage des mod\u00e8les, prolonge la dur\u00e9e de vie de votre fen\u00eatre de contexte et maximise l\u2019efficacit\u00e9 \u00e9conomique de vos caches de prompts.<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter wp-image-170386 size-full\" src=\"https:\/\/jfrog.com\/wp-content\/uploads\/2026\/07\/image1-3.png\" alt=\"\" width=\"960\" height=\"540\" srcset=\"https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2026\/07\/image1-3.png?speedsize=w_960 960w, https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2026\/07\/image1-3.png?speedsize=w_300 300w, https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2026\/07\/image1-3.png?speedsize=w_768 768w\" sizes=\"auto, (max-width: 960px) 100vw, 960px\" \/><\/p>\n<h2>Conception pour une \u00e9conomie d\u2019agents durable<\/h2>\n<p>Le routage de mod\u00e8les permet de r\u00e9aliser de r\u00e9elles \u00e9conomies, mais uniquement lorsqu\u2019il respecte les m\u00e9canismes sous-jacents des caches de prompts des fournisseurs. Pour cr\u00e9er des workflows d\u2019agents durables, traitez votre historique de conversation comme un actif hautement optimis\u00e9&#160;: gardez vos sessions principales l\u00e9g\u00e8res, d\u00e9l\u00e9guez le contexte lourd \u00e0 des sous-agents \u00e9ph\u00e9m\u00e8res et veillez \u00e0 ce que toute transition de mod\u00e8le n\u00e9cessaire soit g\u00e9r\u00e9e au moyen d\u2019un r\u00e9sum\u00e9 d\u2019\u00e9tat structur\u00e9 plut\u00f4t que par une relecture de l\u2019historique brut.<\/p>\n<p>En fin de compte, l\u2019ing\u00e9nierie de syst\u00e8mes agentiques durables exige de traiter l\u2019efficacit\u00e9 des jetons comme une contrainte essentielle, au m\u00eame titre que la vitesse et la pr\u00e9cision.<\/p>\n<p>Pour mettre en \u0153uvre automatiquement ces optimisations au niveau de la session et \u00e9liminer le gaspillage de contexte dans vos workflows de d\u00e9veloppement, vous pouvez <a href=\"http:\/\/boost.jfrog.com\">installer Boost<\/a> et commencer \u00e0 \u00e9conomiser des jetons d\u00e8s aujourd\u2019hui.<\/p>\n<h3>Contactez-nous<\/h3>\n<p>Nous aimerions conna\u00eetre vos exp\u00e9riences, vos sch\u00e9mas d\u2019architecture et vos d\u00e9fis concernant le routage de mod\u00e8les. N\u2019h\u00e9sitez pas \u00e0 nous contacter ou \u00e0 suivre nos recherches en cours sur les agents de codage&#160;:<\/p>\n<p>\u2192 <strong>X&#160;:<\/strong> <a href=\"https:\/\/x.com\/ShayFrektman\" target=\"_blank\" rel=\"noopener\">https:\/\/x.com\/ShayFrektman<\/a>, <a href=\"https:\/\/x.com\/yahav_ohana\" target=\"_blank\" rel=\"noopener\">https:\/\/x.com\/yahav_ohana<\/a><\/p>\n<p>\u2192 <strong>LinkedIn&#160;:<\/strong> <a href=\"https:\/\/www.linkedin.com\/in\/shay-dahan\" target=\"_blank\" rel=\"noopener\">https:\/\/www.linkedin.com\/in\/shay-dahan<\/a>, <a href=\"https:\/\/www.linkedin.com\/in\/yahav-ohana\/\" target=\"_blank\" rel=\"noopener\">https:\/\/www.linkedin.com\/in\/yahav-ohana\/<\/a><\/p>\n<p>\u2192 <strong>Substack&#160;:<\/strong> <a href=\"https:\/\/substack.com\/@yahavohana\" target=\"_blank\" rel=\"noopener\">https:\/\/substack.com\/@yahavohana<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Le routage de mod\u00e8les promet de r\u00e9duire vos d\u00e9penses en agents d\u2019IA en d\u00e9chargeant les t\u00e2ches courantes sur des mod\u00e8les moins chers comme Claude&#160;Haiku, tout en r\u00e9servant les mod\u00e8les de pointe comme Claude&#160;Sonnet au raisonnement complexe. Dans la bonne configuration, les strat\u00e9gies de routage peuvent r\u00e9duire les co\u00fbts d\u2019inf\u00e9rence de 40 \u00e0 85&#160;%. Mais si &hellip;<\/p>\n","protected":false},"author":506,"featured_media":170404,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"inline_featured_image":false,"footnotes":""},"categories":[10619],"tags":[11010,11372,11371,11361,11370,11373],"class_list":["post-174740","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ia-ml","tag-ai-ml-fr","tag-boost","tag-jetons","tag-llm","tag-llm-2","tag-routage-de-modeles","resource_categories-ai","resource_categories-ai-ml"],"yoast_head":"<!-- This site is optimized with the Yoast SEO Premium plugin v27.2 (Yoast SEO v28.6) - https:\/\/yoast.com\/product\/yoast-seo-premium-wordpress\/ -->\n<title>Pourquoi le routage de mod\u00e8les s\u2019av\u00e8re contre-productif et comment cr\u00e9er des agents qui n\u2019\u00e9puisent pas votre budget | JFrog<\/title>\n<meta name=\"description\" content=\"Learn how agentic sessions rapidly accumulate costs, why model routing backfires requires, and what to do instead.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/jfrog.com\/fr\/wp-json\/wp\/v2\/posts\/174740\" \/>\n<meta property=\"og:locale\" content=\"fr_FR\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Pourquoi le routage de mod\u00e8les s\u2019av\u00e8re contre-productif et comment cr\u00e9er des agents qui n\u2019\u00e9puisent pas votre budget\" \/>\n<meta property=\"og:description\" content=\"Learn how agentic sessions rapidly accumulate costs, why model routing backfires requires, and what to do instead.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/jfrog.com\/fr\/blog\/why-model-routing-backfires\/\" \/>\n<meta property=\"og:site_name\" content=\"JFrog\" \/>\n<meta property=\"article:publisher\" content=\"https:\/\/www.facebook.com\/artifrog\" \/>\n<meta property=\"article:published_time\" content=\"2026-07-23T17:16:20+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2026-09-17T04:59:06+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/media.jfrog.com\/wp-content\/uploads\/2026\/07\/23193658\/1200x628-13.png\" \/>\n\t<meta property=\"og:image:width\" content=\"1200\" \/>\n\t<meta property=\"og:image:height\" content=\"628\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/png\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:creator\" content=\"@jfrog\" \/>\n<meta name=\"twitter:site\" content=\"@jfrog\" \/>\n<meta name=\"twitter:label1\" content=\"\u00c9crit par\" \/>\n\t<meta name=\"twitter:data1\" content=\"\" \/>\n\t<meta name=\"twitter:label2\" content=\"Dur\u00e9e de lecture estim\u00e9e\" \/>\n\t<meta name=\"twitter:data2\" content=\"9 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/blog\\\/why-model-routing-backfires\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/blog\\\/why-model-routing-backfires\\\/\"},\"author\":{\"name\":\"\",\"@id\":\"\"},\"headline\":\"Pourquoi le routage de mod\u00e8les s\u2019av\u00e8re contre-productif et comment cr\u00e9er des agents qui n\u2019\u00e9puisent pas votre budget\",\"datePublished\":\"2026-07-23T17:16:20+00:00\",\"dateModified\":\"2026-09-17T04:59:06+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/blog\\\/why-model-routing-backfires\\\/\"},\"wordCount\":1963,\"publisher\":{\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/#organization\"},\"image\":{\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/blog\\\/why-model-routing-backfires\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/speedmedia2.jfrog.com\\\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\\\/jfrog.com\\\/wp-content\\\/uploads\\\/2026\\\/07\\\/Blog_Thumbnail-7-2.png\",\"keywords\":[\"AI\\\/ML\",\"Boost\",\"jetons\",\"LLM\",\"LLM\",\"routage de mod\u00e8les\"],\"articleSection\":[\"IA\\\/ML\"],\"inLanguage\":\"fr-FR\"},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/blog\\\/why-model-routing-backfires\\\/\",\"url\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/blog\\\/why-model-routing-backfires\\\/\",\"name\":\"Pourquoi le routage de mod\u00e8les s\u2019av\u00e8re contre-productif et comment cr\u00e9er des agents qui n\u2019\u00e9puisent pas votre budget | JFrog\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/blog\\\/why-model-routing-backfires\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/blog\\\/why-model-routing-backfires\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/speedmedia2.jfrog.com\\\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\\\/jfrog.com\\\/wp-content\\\/uploads\\\/2026\\\/07\\\/Blog_Thumbnail-7-2.png\",\"datePublished\":\"2026-07-23T17:16:20+00:00\",\"dateModified\":\"2026-09-17T04:59:06+00:00\",\"description\":\"Learn how agentic sessions rapidly accumulate costs, why model routing backfires requires, and what to do instead.\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/blog\\\/why-model-routing-backfires\\\/#breadcrumb\"},\"inLanguage\":\"fr-FR\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/jfrog.com\\\/fr\\\/blog\\\/why-model-routing-backfires\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"fr-FR\",\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/blog\\\/why-model-routing-backfires\\\/#primaryimage\",\"url\":\"https:\\\/\\\/speedmedia2.jfrog.com\\\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\\\/jfrog.com\\\/wp-content\\\/uploads\\\/2026\\\/07\\\/Blog_Thumbnail-7-2.png\",\"contentUrl\":\"https:\\\/\\\/speedmedia2.jfrog.com\\\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\\\/jfrog.com\\\/wp-content\\\/uploads\\\/2026\\\/07\\\/Blog_Thumbnail-7-2.png\",\"width\":203,\"height\":148},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/blog\\\/why-model-routing-backfires\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Pourquoi le routage de mod\u00e8les s\u2019av\u00e8re contre-productif et comment cr\u00e9er des agents qui n\u2019\u00e9puisent pas votre budget\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/#website\",\"url\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/\",\"name\":\"JFrog\",\"description\":\"Deliver Trusted Software Releases at Speed and Scale\",\"publisher\":{\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"fr-FR\"},{\"@type\":\"Organization\",\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/#organization\",\"name\":\"JFrog\",\"url\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"fr-FR\",\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/#\\\/schema\\\/logo\\\/image\\\/\",\"url\":\"https:\\\/\\\/speedmedia2.jfrog.com\\\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\\\/jfrog.com\\\/wp-content\\\/uploads\\\/2025\\\/05\\\/Logo.svg\",\"contentUrl\":\"https:\\\/\\\/speedmedia2.jfrog.com\\\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\\\/jfrog.com\\\/wp-content\\\/uploads\\\/2025\\\/05\\\/Logo.svg\",\"width\":74,\"height\":73,\"caption\":\"JFrog\"},\"image\":{\"@id\":\"https:\\\/\\\/jfrog.com\\\/fr\\\/#\\\/schema\\\/logo\\\/image\\\/\"},\"sameAs\":[\"https:\\\/\\\/www.facebook.com\\\/artifrog\",\"https:\\\/\\\/x.com\\\/jfrog\",\"https:\\\/\\\/www.linkedin.com\\\/company\\\/455737\",\"https:\\\/\\\/www.youtube.com\\\/channel\\\/UCh2hNg76zo3d1qQqTWIQxDg\",\"https:\\\/\\\/www.wikidata.org\\\/wiki\\\/Q98608948\"],\"description\":\"We set out on our Liquid Software journey in 2008, with the mission to transform the way enterprises manage and release software updates. The world expects software to update continuously, securely, non-intrusively and without user intervention. This hyper-connected experience can only be enabled by automation with an end-to-end DevOps platform and a binary-centric focus. With this in mind, we\u2019ve developed the JFrog Platform, ushering in a new era of DevOps and DevSecOps standards that power continuous updates. More than a decade after our founding, with thousands of customers and millions of users globally, JFrog has become the \u201cDatabase of DevOps\u201d and the de-facto standard in release and update management.\",\"legalName\":\"Jfrog, Inc.\",\"numberOfEmployees\":{\"@type\":\"QuantitativeValue\",\"minValue\":\"1001\",\"maxValue\":\"5000\"}},{\"@type\":\"Person\",\"@id\":\"\"}]}<\/script>\n<!-- \/ Yoast SEO Premium plugin. -->","yoast_head_json":{"title":"Pourquoi le routage de mod\u00e8les s\u2019av\u00e8re contre-productif et comment cr\u00e9er des agents qui n\u2019\u00e9puisent pas votre budget | JFrog","description":"Learn how agentic sessions rapidly accumulate costs, why model routing backfires requires, and what to do instead.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/jfrog.com\/fr\/wp-json\/wp\/v2\/posts\/174740","og_locale":"fr_FR","og_type":"article","og_title":"Pourquoi le routage de mod\u00e8les s\u2019av\u00e8re contre-productif et comment cr\u00e9er des agents qui n\u2019\u00e9puisent pas votre budget","og_description":"Learn how agentic sessions rapidly accumulate costs, why model routing backfires requires, and what to do instead.","og_url":"https:\/\/jfrog.com\/fr\/blog\/why-model-routing-backfires\/","og_site_name":"JFrog","article_publisher":"https:\/\/www.facebook.com\/artifrog","article_published_time":"2026-07-23T17:16:20+00:00","article_modified_time":"2026-09-17T04:59:06+00:00","og_image":[{"width":1200,"height":628,"url":"https:\/\/media.jfrog.com\/wp-content\/uploads\/2026\/07\/23193658\/1200x628-13.png","type":"image\/png"}],"twitter_card":"summary_large_image","twitter_creator":"@jfrog","twitter_site":"@jfrog","twitter_misc":{"\u00c9crit par":"","Dur\u00e9e de lecture estim\u00e9e":"9 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/jfrog.com\/fr\/blog\/why-model-routing-backfires\/#article","isPartOf":{"@id":"https:\/\/jfrog.com\/fr\/blog\/why-model-routing-backfires\/"},"author":{"name":"","@id":""},"headline":"Pourquoi le routage de mod\u00e8les s\u2019av\u00e8re contre-productif et comment cr\u00e9er des agents qui n\u2019\u00e9puisent pas votre budget","datePublished":"2026-07-23T17:16:20+00:00","dateModified":"2026-09-17T04:59:06+00:00","mainEntityOfPage":{"@id":"https:\/\/jfrog.com\/fr\/blog\/why-model-routing-backfires\/"},"wordCount":1963,"publisher":{"@id":"https:\/\/jfrog.com\/fr\/#organization"},"image":{"@id":"https:\/\/jfrog.com\/fr\/blog\/why-model-routing-backfires\/#primaryimage"},"thumbnailUrl":"https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2026\/07\/Blog_Thumbnail-7-2.png","keywords":["AI\/ML","Boost","jetons","LLM","LLM","routage de mod\u00e8les"],"articleSection":["IA\/ML"],"inLanguage":"fr-FR"},{"@type":"WebPage","@id":"https:\/\/jfrog.com\/fr\/blog\/why-model-routing-backfires\/","url":"https:\/\/jfrog.com\/fr\/blog\/why-model-routing-backfires\/","name":"Pourquoi le routage de mod\u00e8les s\u2019av\u00e8re contre-productif et comment cr\u00e9er des agents qui n\u2019\u00e9puisent pas votre budget | JFrog","isPartOf":{"@id":"https:\/\/jfrog.com\/fr\/#website"},"primaryImageOfPage":{"@id":"https:\/\/jfrog.com\/fr\/blog\/why-model-routing-backfires\/#primaryimage"},"image":{"@id":"https:\/\/jfrog.com\/fr\/blog\/why-model-routing-backfires\/#primaryimage"},"thumbnailUrl":"https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2026\/07\/Blog_Thumbnail-7-2.png","datePublished":"2026-07-23T17:16:20+00:00","dateModified":"2026-09-17T04:59:06+00:00","description":"Learn how agentic sessions rapidly accumulate costs, why model routing backfires requires, and what to do instead.","breadcrumb":{"@id":"https:\/\/jfrog.com\/fr\/blog\/why-model-routing-backfires\/#breadcrumb"},"inLanguage":"fr-FR","potentialAction":[{"@type":"ReadAction","target":["https:\/\/jfrog.com\/fr\/blog\/why-model-routing-backfires\/"]}]},{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https:\/\/jfrog.com\/fr\/blog\/why-model-routing-backfires\/#primaryimage","url":"https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2026\/07\/Blog_Thumbnail-7-2.png","contentUrl":"https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2026\/07\/Blog_Thumbnail-7-2.png","width":203,"height":148},{"@type":"BreadcrumbList","@id":"https:\/\/jfrog.com\/fr\/blog\/why-model-routing-backfires\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/jfrog.com\/fr\/"},{"@type":"ListItem","position":2,"name":"Pourquoi le routage de mod\u00e8les s\u2019av\u00e8re contre-productif et comment cr\u00e9er des agents qui n\u2019\u00e9puisent pas votre budget"}]},{"@type":"WebSite","@id":"https:\/\/jfrog.com\/fr\/#website","url":"https:\/\/jfrog.com\/fr\/","name":"JFrog","description":"Deliver Trusted Software Releases at Speed and Scale","publisher":{"@id":"https:\/\/jfrog.com\/fr\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/jfrog.com\/fr\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"fr-FR"},{"@type":"Organization","@id":"https:\/\/jfrog.com\/fr\/#organization","name":"JFrog","url":"https:\/\/jfrog.com\/fr\/","logo":{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https:\/\/jfrog.com\/fr\/#\/schema\/logo\/image\/","url":"https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2025\/05\/Logo.svg","contentUrl":"https:\/\/speedmedia2.jfrog.com\/08612fe1-9391-4cf3-ac1a-6dd49c36b276\/jfrog.com\/wp-content\/uploads\/2025\/05\/Logo.svg","width":74,"height":73,"caption":"JFrog"},"image":{"@id":"https:\/\/jfrog.com\/fr\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/www.facebook.com\/artifrog","https:\/\/x.com\/jfrog","https:\/\/www.linkedin.com\/company\/455737","https:\/\/www.youtube.com\/channel\/UCh2hNg76zo3d1qQqTWIQxDg","https:\/\/www.wikidata.org\/wiki\/Q98608948"],"description":"We set out on our Liquid Software journey in 2008, with the mission to transform the way enterprises manage and release software updates. The world expects software to update continuously, securely, non-intrusively and without user intervention. This hyper-connected experience can only be enabled by automation with an end-to-end DevOps platform and a binary-centric focus. With this in mind, we\u2019ve developed the JFrog Platform, ushering in a new era of DevOps and DevSecOps standards that power continuous updates. More than a decade after our founding, with thousands of customers and millions of users globally, JFrog has become the \u201cDatabase of DevOps\u201d and the de-facto standard in release and update management.","legalName":"Jfrog, Inc.","numberOfEmployees":{"@type":"QuantitativeValue","minValue":"1001","maxValue":"5000"}},{"@type":"Person","@id":""}]}},"_links":{"self":[{"href":"https:\/\/jfrog.com\/fr\/wp-json\/wp\/v2\/posts\/174740","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/jfrog.com\/fr\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/jfrog.com\/fr\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/jfrog.com\/fr\/wp-json\/wp\/v2\/users\/506"}],"replies":[{"embeddable":true,"href":"https:\/\/jfrog.com\/fr\/wp-json\/wp\/v2\/comments?post=174740"}],"version-history":[{"count":0,"href":"https:\/\/jfrog.com\/fr\/wp-json\/wp\/v2\/posts\/174740\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/jfrog.com\/fr\/wp-json\/wp\/v2\/media\/170404"}],"wp:attachment":[{"href":"https:\/\/jfrog.com\/fr\/wp-json\/wp\/v2\/media?parent=174740"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/jfrog.com\/fr\/wp-json\/wp\/v2\/categories?post=174740"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/jfrog.com\/fr\/wp-json\/wp\/v2\/tags?post=174740"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}