[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-post-\u002Ffr\u002Fnews\u002Fopenai-gpt-55-super-app-codex-atlas-agent-avril-2026":3,"news-related-fr-\u002Ffr\u002Fnews\u002Fopenai-gpt-55-super-app-codex-atlas-agent-avril-2026":640,"hreflang-openai-gpt-55-super-app-codex-atlas-agent":660},{"doc":4,"redirect":8},{"id":5,"title":6,"author":7,"author_avatar":8,"body":9,"category":615,"date":616,"description":617,"extension":618,"image":619,"meta":620,"navigation":621,"path":622,"seo":623,"seoDescription":624,"seoTitle":625,"stem":626,"tags":627,"translationKey":638,"updatedAt":8,"__hash__":639},"news_fr\u002Ffr\u002Fnews\u002Fopenai-gpt-55-super-app-codex-atlas-agent-avril-2026.md","OpenAI lance GPT-5.5 : le modèle agentique qui pose la première pierre de la super app ChatGPT + Codex + Atlas","Paul Lefizelier",null,{"type":10,"value":11,"toc":604},"minimark",[12,45,50,58,83,97,101,104,191,204,208,215,277,285,289,292,298,308,314,318,325,333,337,456,464,468,471,477,488,499,502,509,551,553,556,559],[13,14,15,16,20,21,24,25,28,29,32,33,36,37,40,41,44],"p",{},"OpenAI n'a plus envie de vendre un assistant. Le 23 avril 2026, la société a dévoilé ",[17,18,19],"strong",{},"GPT-5.5",", présenté par Sam Altman comme « le modèle le plus intelligent et le plus intuitif » de la maison, et par Greg Brockman comme « la première brique concrète de la super app ». Sous le capot, c'est le premier modèle ",[17,22,23],{},"entièrement réentraîné"," depuis GPT-4.5 — pas un simple post-training appliqué à GPT-5. Et les scores sont agressifs : ",[17,26,27],{},"84,9 %"," sur GDPval, ",[17,30,31],{},"78,7 %"," sur OSWorld-Verified, ",[17,34,35],{},"82,7 %"," sur Terminal-Bench 2.0, ",[17,38,39],{},"98,0 %"," sur Tau2-bench Telecom. Le tout sur un prix API doublé — ",[17,42,43],{},"5 $ en entrée, 30 $ en sortie"," par million de tokens, contre 2,50 $ \u002F 15 $ pour GPT-5.4. Le message est clair : le token moins cher n'est plus le bon champ de bataille.",[46,47,49],"h2",{"id":48},"le-pari-super-app-une-seule-session-plusieurs-surfaces","Le pari super app : une seule session, plusieurs surfaces",[13,51,52,53,57],{},"Depuis trois mois, OpenAI prépare quelque chose qui ne ressemble plus à une API. La stratégie, nommée en interne ",[54,55,56],"em",{},"unified desktop",", fusionne dans une même session :",[59,60,61,68,77],"ul",{},[62,63,64,67],"li",{},[17,65,66],{},"ChatGPT"," — la conversation principale et la mémoire utilisateur",[62,69,70,73,74],{},[17,71,72],{},"Codex"," — l'environnement de développement et la couche ",[54,75,76],{},"computer use",[62,78,79,82],{},[17,80,81],{},"Atlas"," — le navigateur-agent qui exécute dans le DOM",[13,84,85,86,91,92,96],{},"GPT-5.5 est le modèle sous-jacent unique qui coud les trois. Quand l'utilisateur passe d'une tâche de recherche dans ChatGPT à un déploiement dans Codex à une navigation dans Atlas, le contexte n'est pas rechargé — il reste dans la même fenêtre mentale du modèle. Le parallèle évident est ",[87,88,90],"a",{"href":89},"\u002Ffr\u002Fnews\u002Fclaude-computer-use-cowork-mac-controle-total-agent-ia","Claude Code dans le Mac que Claude Computer Use a annoncé l'automne dernier",", mais avec le front-end grand public d'OpenAI derrière. C'est aussi la réponse directe à ",[87,93,95],{"href":94},"\u002Ffr\u002Fnews\u002Fopenai-codex-desktop-control-ordinateur-mac-90-plugins-memoire-avril-2026","Codex Desktop, annoncé il y a trois jours avec 90 plugins et de la mémoire"," — GPT-5.5 est le cerveau que Codex Desktop attendait.",[46,98,100],{"id":99},"les-benchmarks-et-ce-quils-révèlent-du-positionnement","Les benchmarks — et ce qu'ils révèlent du positionnement",[13,102,103],{},"Les scores publiés par OpenAI racontent une histoire précise : on arrête de courir derrière SWE-Bench pure et on se positionne sur le travail réel.",[105,106,107,123],"table",{},[108,109,110],"thead",{},[111,112,113,117,120],"tr",{},[114,115,116],"th",{},"Benchmark",[114,118,119],{},"Score GPT-5.5",[114,121,122],{},"Ce qui est mesuré",[124,125,126,139,151,163,176],"tbody",{},[111,127,128,134,136],{},[129,130,131],"td",{},[17,132,133],{},"GDPval",[129,135,27],{},[129,137,138],{},"Qualité des livrables sur 44 métiers du knowledge work",[111,140,141,146,148],{},[129,142,143],{},[17,144,145],{},"OSWorld-Verified",[129,147,31],{},[129,149,150],{},"Exécution autonome dans de vrais environnements OS",[111,152,153,158,160],{},[129,154,155],{},[17,156,157],{},"Terminal-Bench 2.0",[129,159,35],{},[129,161,162],{},"Capacité à conduire des sessions shell longues",[111,164,165,170,173],{},[129,166,167],{},[17,168,169],{},"SWE-Bench Pro",[129,171,172],{},"58,6 %",[129,174,175],{},"Ingénierie logicielle sur tâches réelles",[111,177,178,183,185],{},[129,179,180],{},[17,181,182],{},"Tau2-bench Telecom",[129,184,39],{},[129,186,187,188],{},"Agents métier sans ",[54,189,190],{},"prompt tuning",[13,192,193,194,198,199,203],{},"Le choix de ces benchmarks est délibéré. GDPval, OSWorld et Terminal-Bench mesurent l'agent — pas la complétion de snippets. Sur SWE-Bench Pro, 58,6 %, c'est honorable mais pas le podium. Pour rappel, ",[87,195,197],{"href":196},"\u002Ffr\u002Fnews\u002Fanthropic-claude-opus-4-7-swe-bench-87-mythos-avril-2026","Claude Opus 4.7 est publié à 87 % sur SWE-Bench",", et ",[87,200,202],{"href":201},"\u002Ffr\u002Fnews\u002Falibaba-qwen-36-35b-a3b-moe-open-source-apache-swe-bench-avril-2026","Qwen 3.6 35B-A3B approche les 75 % sur la même épreuve",". OpenAI accepte de ne plus être premier sur l'écriture de code pure et revendique la première place sur « l'ordinateur comme terrain d'action ». C'est un choix de marché : Anthropic pour les IDE, OpenAI pour l'écran.",[46,205,207],{"id":206},"le-prix-doublé-et-ce-que-ça-dit-du-marché","Le prix doublé — et ce que ça dit du marché",[13,209,210,211,214],{},"La grosse surprise du lancement n'est pas technique. Elle est commerciale. GPT-5.5 est ",[17,212,213],{},"deux fois plus cher à l'API"," que GPT-5.4 :",[105,216,217,230],{},[108,218,219],{},[111,220,221,224,227],{},[114,222,223],{},"Modèle",[114,225,226],{},"Input ($\u002FM tokens)",[114,228,229],{},"Output ($\u002FM tokens)",[124,231,232,245,261],{},[111,233,234,239,242],{},[129,235,236],{},[17,237,238],{},"GPT-5.4",[129,240,241],{},"2,50",[129,243,244],{},"15,00",[111,246,247,251,256],{},[129,248,249],{},[17,250,19],{},[129,252,253],{},[17,254,255],{},"5,00",[129,257,258],{},[17,259,260],{},"30,00",[111,262,263,268,272],{},[129,264,265],{},[17,266,267],{},"GPT-5.5 Pro",[129,269,270],{},[17,271,260],{},[129,273,274],{},[17,275,276],{},"180,00",[13,278,279,280,284],{},"Dans un marché où Qwen 3.6 est open-source sous Apache 2.0 et où Gemini 2.5 Flash descend en dessous du dollar par million d'input tokens, doubler le prix est un pari. OpenAI le justifie par une efficacité de raisonnement supérieure — « moins de tokens pour une meilleure réponse » — et par la valeur créée pour les cas d'usage agentiques. Autrement dit : le prix au token devient un mauvais signal, ce qui compte c'est le prix à la tâche accomplie. C'est la même ligne éditoriale qu'",[87,281,283],{"href":282},"\u002Ffr\u002Fnews\u002Fanthropic-refuse-800-milliards-valorisation-vc-offres-preemptives-avril-2026","Anthropic suit en refusant la valorisation à 800 milliards pour préserver sa trajectoire prix"," — personne n'a envie d'être Uber 2015 en 2026.",[46,286,288],{"id":287},"mémoire-outils-gpt-55-pro","Mémoire, outils, GPT-5.5 Pro",[13,290,291],{},"Trois éléments sous-estimés du lancement méritent d'être listés :",[13,293,294,297],{},[17,295,296],{},"La mémoire de session étendue."," GPT-5.5 garde en contexte une session de plusieurs heures sans resumer agressif. C'est le prérequis pour qu'un agent de programmation ne redemande pas « dans quel fichier déjà ? » après quatre appels d'outils.",[13,299,300,303,304,307],{},[17,301,302],{},"Le switch d'outils piloté par le modèle."," Plutôt que de laisser l'orchestrateur appeler les outils en boucle externe, GPT-5.5 décide lui-même quand appeler l'",[54,305,306],{},"atlas",", le terminal, l'éditeur, ou le file system. L'orchestrateur devient une garde-fou, plus un routeur.",[13,309,310,313],{},[17,311,312],{},"GPT-5.5 Pro — réservé aux tâches longues."," Le tier Pro à 30 $ \u002F 180 $ est positionné sur les recherches scientifiques, les rapports multi-étapes, les audits. C'est le seul modèle du line-up à pouvoir maintenir une chaîne de raisonnement de plus de douze outils sans perte de cohérence, d'après les chiffres internes d'OpenAI.",[46,315,317],{"id":316},"disponibilité-et-déploiement","Disponibilité et déploiement",[13,319,320,321,324],{},"Le modèle est ",[17,322,323],{},"disponible depuis le 23 avril 2026"," sur les tiers Plus, Pro, Business et Enterprise de ChatGPT. La variante Pro est réservée aux tiers Pro, Business et Enterprise. L'API est ouverte dès le jour du lancement, sans liste d'attente.",[13,326,327,328,332],{},"Côté infrastructure, GPT-5.5 tourne sur les clusters Nvidia que ",[87,329,331],{"href":330},"\u002Ffr\u002Fnews\u002Fneocognition-40-millions-seed-agents-ia-auto-apprenants-2026","NeoCognition et d'autres sont aussi en train de remplir"," — la pénurie de GPU n'est toujours pas résolue, mais OpenAI a sécurisé assez de capacité pour un déploiement mondial le jour 1. C'est la seule manière de défendre un prix doublé : rendre la disponibilité un argument.",[46,334,336],{"id":335},"le-positionnement-vs-claude-gemini-qwen","Le positionnement vs Claude, Gemini, Qwen",[105,338,339,357],{},[108,340,341],{},[111,342,343,346,348,351,354],{},[114,344,345],{},"Axe",[114,347,19],{},[114,349,350],{},"Claude Opus 4.7",[114,352,353],{},"Gemini 2.5 Pro",[114,355,356],{},"Qwen 3.6",[124,358,359,380,401,420,437],{},[111,360,361,366,369,374,377],{},[129,362,363],{},[17,364,365],{},"SWE-Bench",[129,367,368],{},"58,6 % (Pro)",[129,370,371],{},[17,372,373],{},"87 %",[129,375,376],{},"~72 %",[129,378,379],{},"~75 %",[111,381,382,387,392,395,398],{},[129,383,384],{},[17,385,386],{},"Computer Use",[129,388,389],{},[17,390,391],{},"78,7 % OSWorld",[129,393,394],{},"Bon",[129,396,397],{},"Project Mariner",[129,399,400],{},"Limité",[111,402,403,408,411,414,417],{},[129,404,405],{},[17,406,407],{},"Prix input",[129,409,410],{},"5,00 $",[129,412,413],{},"3,00 $",[129,415,416],{},"1,25 $",[129,418,419],{},"Open-source",[111,421,422,427,430,432,434],{},[129,423,424],{},[17,425,426],{},"Licence",[129,428,429],{},"Propriétaire",[129,431,429],{},[129,433,429],{},[129,435,436],{},"Apache 2.0",[111,438,439,444,447,450,453],{},[129,440,441],{},[17,442,443],{},"Stratégie",[129,445,446],{},"Super app consumer",[129,448,449],{},"API + enterprise",[129,451,452],{},"Full-stack cloud",[129,454,455],{},"Open-weight",[13,457,458,459,463],{},"OpenAI choisit d'être le modèle de la surface grand public. Claude reste le modèle des IDE. Gemini est le modèle du cloud enterprise, comme l'a confirmé ",[87,460,462],{"href":461},"\u002Ffr\u002Fnews\u002Fgoogle-cloud-next-2026-gemini-enterprise-agent-platform-a2a-tpu8-avril-2026","Google Cloud Next 2026 avec Gemini Enterprise Agent Platform et le protocole A2A en production chez 150 organisations",". Et Qwen est le modèle des équipes qui refusent de payer une API. Le paysage se segmente enfin par cas d'usage, pas par « qui a le plus gros benchmark ».",[46,465,467],{"id":466},"ce-que-ça-signifie-pour-les-développeurs-et-les-publishers","Ce que ça signifie pour les développeurs et les publishers",[13,469,470],{},"Pour les développeurs qui construisent des apps IA, trois conséquences :",[13,472,473,476],{},[17,474,475],{},"Le prompt engineering pour l'agent devient un métier."," GPT-5.5 excelle sur les tâches mal spécifiées mais s'effondre sur les orchestrations mal décrites. L'écart entre un agent qui livre et un agent qui tourne en rond se creuse.",[13,478,479,482,483,487],{},[17,480,481],{},"Le prix au token cesse d'être le KPI pertinent."," Les équipes produit vont tracker le « prix à la tâche accomplie », qui inclut les retries, le contexte, la mémoire. Celles qui l'ont déjà compris sont celles qui ",[87,484,486],{"href":485},"\u002Ffr\u002Fpublishers","monétisent leurs apps IA avec @idlen\u002Fchat-sdk — où chaque token inutilisé par l'utilisateur ne disparaît pas, il capitalise",".",[13,489,490,493,494,498],{},[17,491,492],{},"L'agent devient l'unité d'abonnement."," ChatGPT Pro à 200 $\u002Fmois avec GPT-5.5 Pro illimité n'est pas un prix, c'est un abonnement à la production d'un agent permanent. ",[87,495,497],{"href":496},"\u002Ffr\u002Fnews\u002Fcursor-3-interface-agents-parallele-vibe-coding-2026","Lovable, Cursor et Emergent suivent déjà cette ligne",". OpenAI vient de la normaliser côté grand public.",[500,501],"hr",{},[13,503,504],{},[54,505,506],{},[17,507,508],{},"En résumé :",[59,510,511,521,527,533,539,545],{},[62,512,513,516,517,520],{},[17,514,515],{},"OpenAI a lancé GPT-5.5"," le ",[17,518,519],{},"23 avril 2026"," — premier modèle entièrement réentraîné depuis GPT-4.5.",[62,522,523,526],{},[17,524,525],{},"Benchmarks clés"," : 84,9 % sur GDPval, 78,7 % sur OSWorld-Verified, 82,7 % sur Terminal-Bench 2.0, 98,0 % sur Tau2-bench Telecom, 58,6 % sur SWE-Bench Pro.",[62,528,529,532],{},[17,530,531],{},"Prix API doublé"," : 5 $ \u002F 30 $ par million de tokens (10 $ \u002F 60 $ pour Pro), avec la variante GPT-5.5 Pro à 30 $ \u002F 180 $.",[62,534,535,538],{},[17,536,537],{},"La super app"," fusionne ChatGPT, Codex et Atlas sous un seul modèle et une seule session.",[62,540,541,544],{},[17,542,543],{},"Disponible dès le jour 1"," sur Plus, Pro, Business, Enterprise et l'API.",[62,546,547,550],{},[17,548,549],{},"Positionnement"," : OpenAI concède SWE-Bench à Claude et revendique la première place sur l'ordinateur comme terrain d'action.",[500,552],{},[13,554,555],{},"GPT-5.5 est un pari sur la forme, pas sur le modèle. OpenAI a compris que l'intelligence brute ne gagne plus de parts de marché — c'est l'intégration dans la vie quotidienne du développeur, du knowledge worker, du chercheur qui fait la différence. La super app ChatGPT + Codex + Atlas est l'incarnation de ce pari. Reste à voir si les entreprises acceptent de payer deux fois le prix pour un modèle qui fait 58 % sur SWE-Bench quand Claude Opus 4.7 en fait 87 %. La réponse arrivera avec les chiffres de revenus du Q2 2026.",[13,557,558],{},"Sources :",[59,560,561,569,576,583,590,597],{},[62,562,563],{},[87,564,568],{"href":565,"rel":566},"https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-gpt-5-5\u002F",[567],"nofollow","Introducing GPT-5.5 — OpenAI",[62,570,571],{},[87,572,575],{"href":573,"rel":574},"https:\u002F\u002Ftechcrunch.com\u002F2026\u002F04\u002F23\u002Fopenai-chatgpt-gpt-5-5-ai-model-superapp\u002F",[567],"OpenAI releases GPT-5.5, bringing company one step closer to an AI 'super app' — TechCrunch",[62,577,578],{},[87,579,582],{"href":580,"rel":581},"https:\u002F\u002Fblogs.nvidia.com\u002Fblog\u002Fopenai-codex-gpt-5-5-ai-agents\u002F",[567],"OpenAI's New GPT-5.5 Powers Codex on NVIDIA Infrastructure — NVIDIA Blog",[62,584,585],{},[87,586,589],{"href":587,"rel":588},"https:\u002F\u002Fthenextweb.com\u002Fnews\u002Fopenai-gpt-5-5-launch-enterprise",[567],"OpenAI launches GPT-5.5, its first fully retrained base model since GPT-4.5 — The Next Web",[62,591,592],{},[87,593,596],{"href":594,"rel":595},"https:\u002F\u002Fthe-decoder.com\u002Fopenai-unveils-gpt-5-5-claims-a-new-class-of-intelligence-at-double-the-api-price\u002F",[567],"OpenAI unveils GPT-5.5, claims a \"new class of intelligence\" at double the API price — The Decoder",[62,598,599],{},[87,600,603],{"href":601,"rel":602},"https:\u002F\u002Fapidog.com\u002Fblog\u002Fgpt-5-5-pricing\u002F",[567],"GPT-5.5 Pricing: Full Breakdown — Apidog",{"title":605,"searchDepth":606,"depth":606,"links":607},"",2,[608,609,610,611,612,613,614],{"id":48,"depth":606,"text":49},{"id":99,"depth":606,"text":100},{"id":206,"depth":606,"text":207},{"id":287,"depth":606,"text":288},{"id":316,"depth":606,"text":317},{"id":335,"depth":606,"text":336},{"id":466,"depth":606,"text":467},"ai","2026-04-24T09:00:00","Le 23 avril 2026, OpenAI a dévoilé GPT-5.5, son premier modèle entièrement réentraîné depuis GPT-4.5. 84,9 % sur GDPval, 78,7 % sur OSWorld, 82,7 % sur Terminal-Bench 2.0 — et un prix API doublé à 5 $ \u002F 30 $ par million de tokens. Objectif : fusionner ChatGPT, Codex et Atlas en une seule session agentique.","md","\u002Fopenai-gpt-55-super-app-codex-atlas-agent.jpg",{},true,"\u002Ffr\u002Fnews\u002Fopenai-gpt-55-super-app-codex-atlas-agent-avril-2026",{"title":6,"description":617},"Le 23 avril 2026, OpenAI lance GPT-5.5 — 84,9 % GDPval, 78,7 % OSWorld, 82,7 % Terminal-Bench 2.0. Premier modèle entièrement réentraîné depuis GPT-4.5. API à 5 $ \u002F 30 $ par million de tokens. La super app ChatGPT + Codex + Atlas prend forme.","OpenAI GPT-5.5 : super app agentique ChatGPT + Codex + Atlas — avril 2026","fr\u002Fnews\u002Fopenai-gpt-55-super-app-codex-atlas-agent-avril-2026",[628,629,630,631,632,633,634,635,636,637],"openai","gpt-5-5","chatgpt","codex","atlas-browser","super-app","agentic-ai","agent","computer-use","gpt-5","openai-gpt-55-super-app-codex-atlas-agent","7SSpfUtp-yykkEap2z6zbKQ4BjYok6YYzOvGBIWPshU",[641,648,654],{"title":642,"description":643,"path":644,"date":645,"image":646,"category":647},"Les cloud agents Cursor construisent, hébergent et publient des apps live","Les cloud agents Cursor peuvent scaffolder une app web, provisionner Postgres et le SSO Cursor, lancer des contrôles de sécurité, puis publier une URL partageable — sans compte d'hébergement séparé.","\u002Ffr\u002Fnews\u002Fcursor-live-apps-agents-cloud-deploiement-postgres-auth","2026-11-07T08:00:00","\u002Fimages\u002Fnews\u002Fcursor-live-apps-cloud-agents-deploy-postgres-auth.jpg","vibecoding",{"title":649,"description":650,"path":651,"date":652,"image":653,"category":615},"Google rend Gemini Omni 1.2 Flash généralement disponible, avec des vidéos de 15 secondes","Google DeepMind baisse le prix d'Omni Flash de 30 %, à 0,07 $ par seconde en 720p, passe de 10 à 15 secondes de clip, et ouvre le modèle sur l'API Gemini et Google Cloud.","\u002Ffr\u002Fnews\u002Fgoogle-gemini-omni-1-2-flash-video-15-secondes","2026-11-06T08:00:00","\u002Fimages\u002Fnews\u002Fgoogle-gemini-omni-1-2-flash-15-second-video.jpg",{"title":655,"description":656,"path":657,"date":658,"image":659,"category":615},"Anthropic aurait signé son premier accord de licence avec un éditeur, l'AP","Selon Adweek, Anthropic a conclu un accord de licence de contenus et de données avec l'Associated Press juste avant les élections de mi-mandat américaines, son premier avec un éditeur numérique. Aucune des deux entreprises ne l'a confirmé officiellement, et les conditions ne sont pas publiques.","\u002Ffr\u002Fnews\u002Fanthropic-premier-accord-licence-editeurs-associated-press","2026-11-05T12:35:00","\u002Fimages\u002Fnews\u002Fanthropic-ap-publisher-licensing-deal.jpg",{"en":661,"fr":1223,"es":8},{"id":662,"title":663,"author":7,"author_avatar":8,"body":664,"category":615,"date":616,"description":1214,"extension":618,"image":619,"meta":1215,"navigation":621,"path":1216,"seo":1217,"seoDescription":1218,"seoTitle":1219,"stem":1220,"tags":1221,"translationKey":638,"updatedAt":8,"__hash__":1222},"news_en\u002Fnews\u002Fopenai-gpt-55-super-app-codex-atlas-agent-april-2026.md","OpenAI Launches GPT-5.5: The Agentic Model Laying the First Brick of the ChatGPT + Codex + Atlas Super App",{"type":10,"value":665,"toc":1205},[666,700,704,710,730,743,747,750,824,837,841,848,906,914,918,921,927,933,939,943,950,958,962,1072,1080,1084,1087,1093,1103,1114,1116,1123,1165,1167,1170,1173],[13,667,668,669,671,672,675,676,679,680,683,684,687,688,691,692,695,696,699],{},"OpenAI is no longer in the business of selling an assistant. On April 23, 2026, the company unveiled ",[17,670,19],{},", which Sam Altman called \"the most intelligent and most intuitive-to-use model\" the house has shipped, and which Greg Brockman described as \"the first concrete brick of the super app.\" Under the hood, it's the ",[17,673,674],{},"first fully retrained base model"," since GPT-4.5 — not post-training layered on top of GPT-5. And the scores are aggressive: ",[17,677,678],{},"84.9%"," on GDPval, ",[17,681,682],{},"78.7%"," on OSWorld-Verified, ",[17,685,686],{},"82.7%"," on Terminal-Bench 2.0, ",[17,689,690],{},"98.0%"," on Tau2-bench Telecom. All of it on ",[17,693,694],{},"doubled API pricing"," — ",[17,697,698],{},"$5 input, $30 output"," per million tokens, against $2.50 \u002F $15 for GPT-5.4. The message is clear: the cheap token is no longer the battlefield.",[46,701,703],{"id":702},"the-super-app-bet-one-session-multiple-surfaces","The super app bet: one session, multiple surfaces",[13,705,706,707,709],{},"For the past three months, OpenAI has been quietly engineering something that doesn't look like an API anymore. The strategy, dubbed ",[54,708,56],{}," internally, merges into a single session:",[59,711,712,717,725],{},[62,713,714,716],{},[17,715,66],{}," — the primary conversation and user memory",[62,718,719,721,722,724],{},[17,720,72],{}," — the development environment and the ",[54,723,76],{}," surface",[62,726,727,729],{},[17,728,81],{}," — the browser-agent executing against the DOM",[13,731,732,733,737,738,742],{},"GPT-5.5 is the single underlying model stitching the three together. When a user moves from research in ChatGPT to a deploy in Codex to browsing in Atlas, the context isn't reloaded — it stays inside the same mental window of the model. The obvious parallel is ",[87,734,736],{"href":735},"\u002Fnews\u002Fclaude-computer-use-cowork-mac-full-control-ai-agent","Claude Code on Mac announced last fall with Claude Computer Use",", but with OpenAI's consumer front end behind it. It's also the direct answer to ",[87,739,741],{"href":740},"\u002Fnews\u002Fopenai-codex-desktop-computer-use-mac-90-plugins-memory-april-2026","Codex Desktop, announced three days ago with 90 plugins and memory"," — GPT-5.5 is the brain Codex Desktop was waiting for.",[46,744,746],{"id":745},"the-benchmarks-and-what-they-reveal-about-positioning","The benchmarks — and what they reveal about positioning",[13,748,749],{},"The numbers OpenAI published tell a specific story: we're done chasing raw SWE-Bench and we're positioning on real work.",[105,751,752,764],{},[108,753,754],{},[111,755,756,758,761],{},[114,757,116],{},[114,759,760],{},"GPT-5.5 score",[114,762,763],{},"What it measures",[124,765,766,777,788,799,811],{},[111,767,768,772,774],{},[129,769,770],{},[17,771,133],{},[129,773,678],{},[129,775,776],{},"Quality of deliverables across 44 knowledge-work occupations",[111,778,779,783,785],{},[129,780,781],{},[17,782,145],{},[129,784,682],{},[129,786,787],{},"Autonomous execution in real OS environments",[111,789,790,794,796],{},[129,791,792],{},[17,793,157],{},[129,795,686],{},[129,797,798],{},"Ability to drive long shell sessions",[111,800,801,805,808],{},[129,802,803],{},[17,804,169],{},[129,806,807],{},"58.6%",[129,809,810],{},"Real-world software engineering tasks",[111,812,813,817,819],{},[129,814,815],{},[17,816,182],{},[129,818,690],{},[129,820,821,822],{},"Business agents without ",[54,823,190],{},[13,825,826,827,831,832,836],{},"The benchmark selection is deliberate. GDPval, OSWorld, and Terminal-Bench measure the agent — not snippet completion. On SWE-Bench Pro, 58.6% is respectable but not the podium. For reference, ",[87,828,830],{"href":829},"\u002Fnews\u002Fanthropic-claude-opus-4-7-swe-bench-87-mythos-april-2026","Claude Opus 4.7 ships at 87% on SWE-Bench",", and ",[87,833,835],{"href":834},"\u002Fnews\u002Falibaba-qwen-36-35b-a3b-moe-open-source-apache-swe-bench-april-2026","Qwen 3.6 35B-A3B hits around 75% on the same eval",". OpenAI is openly conceding the raw-coding title and claiming pole position on \"the computer as the agent's playing field.\" It's a market call: Anthropic for the IDE, OpenAI for the screen.",[46,838,840],{"id":839},"doubled-pricing-and-what-it-says-about-the-market","Doubled pricing — and what it says about the market",[13,842,843,844,847],{},"The big surprise of the launch isn't technical. It's commercial. GPT-5.5 is ",[17,845,846],{},"twice as expensive on the API"," as GPT-5.4:",[105,849,850,861],{},[108,851,852],{},[111,853,854,857,859],{},[114,855,856],{},"Model",[114,858,226],{},[114,860,229],{},[124,862,863,875,891],{},[111,864,865,869,872],{},[129,866,867],{},[17,868,238],{},[129,870,871],{},"2.50",[129,873,874],{},"15.00",[111,876,877,881,886],{},[129,878,879],{},[17,880,19],{},[129,882,883],{},[17,884,885],{},"5.00",[129,887,888],{},[17,889,890],{},"30.00",[111,892,893,897,901],{},[129,894,895],{},[17,896,267],{},[129,898,899],{},[17,900,890],{},[129,902,903],{},[17,904,905],{},"180.00",[13,907,908,909,913],{},"In a market where Qwen 3.6 is open-source under Apache 2.0 and where Gemini 2.5 Flash drops below a dollar per million input tokens, doubling the price is a bet. OpenAI justifies it on superior reasoning efficiency — \"fewer tokens for a better answer\" — and on the value created for agentic use cases. In other words: the per-token price is the wrong signal, what matters is the price per task completed. It's the same editorial line that ",[87,910,912],{"href":911},"\u002Fnews\u002Fanthropic-refuses-800-billion-valuation-vc-preemptive-offers-april-2026","Anthropic holds by refusing the $800 billion valuation to preserve its pricing trajectory"," — no one wants to be Uber-2015 in 2026.",[46,915,917],{"id":916},"memory-tools-gpt-55-pro","Memory, tools, GPT-5.5 Pro",[13,919,920],{},"Three underrated elements of the launch worth listing:",[13,922,923,926],{},[17,924,925],{},"Extended session memory."," GPT-5.5 holds a multi-hour session in context without aggressive resummarization. That's the prerequisite so a programming agent doesn't ask \"which file was it again?\" after four tool calls.",[13,928,929,932],{},[17,930,931],{},"Model-driven tool switching."," Rather than letting the orchestrator call tools in an outer loop, GPT-5.5 itself decides when to call the atlas, the terminal, the editor, or the file system. The orchestrator becomes a guardrail, not a router.",[13,934,935,938],{},[17,936,937],{},"GPT-5.5 Pro — reserved for long tasks."," The Pro tier at $30 \u002F $180 is positioned for scientific research, multi-step reports, audits. It's the only model in the line-up capable of maintaining a chain of reasoning across more than twelve tools without losing coherence, per OpenAI's internal figures.",[46,940,942],{"id":941},"availability-and-deployment","Availability and deployment",[13,944,945,946,949],{},"The model has been ",[17,947,948],{},"available since April 23, 2026"," on the Plus, Pro, Business, and Enterprise tiers of ChatGPT. The Pro variant is reserved for Pro, Business, and Enterprise tiers. The API opened on launch day with no waitlist.",[13,951,952,953,957],{},"On the infrastructure side, GPT-5.5 runs on the Nvidia clusters ",[87,954,956],{"href":955},"\u002Fnews\u002Fneocognition-40-million-seed-self-learning-ai-agents-2026","that NeoCognition and others are also filling up"," — the GPU shortage isn't over, but OpenAI has locked in enough capacity for a global day-one rollout. It's the only way to defend doubled pricing: make availability an argument.",[46,959,961],{"id":960},"positioning-vs-claude-gemini-qwen","Positioning vs Claude, Gemini, Qwen",[105,963,964,979],{},[108,965,966],{},[111,967,968,971,973,975,977],{},[114,969,970],{},"Axis",[114,972,19],{},[114,974,350],{},[114,976,353],{},[114,978,356],{},[124,980,981,1001,1021,1040,1056],{},[111,982,983,987,990,995,998],{},[129,984,985],{},[17,986,365],{},[129,988,989],{},"58.6% (Pro)",[129,991,992],{},[17,993,994],{},"87%",[129,996,997],{},"~72%",[129,999,1000],{},"~75%",[111,1002,1003,1008,1013,1016,1018],{},[129,1004,1005],{},[17,1006,1007],{},"Computer use",[129,1009,1010],{},[17,1011,1012],{},"78.7% OSWorld",[129,1014,1015],{},"Strong",[129,1017,397],{},[129,1019,1020],{},"Limited",[111,1022,1023,1028,1031,1034,1037],{},[129,1024,1025],{},[17,1026,1027],{},"Input price",[129,1029,1030],{},"$5.00",[129,1032,1033],{},"$3.00",[129,1035,1036],{},"$1.25",[129,1038,1039],{},"Open source",[111,1041,1042,1047,1050,1052,1054],{},[129,1043,1044],{},[17,1045,1046],{},"License",[129,1048,1049],{},"Proprietary",[129,1051,1049],{},[129,1053,1049],{},[129,1055,436],{},[111,1057,1058,1063,1066,1068,1070],{},[129,1059,1060],{},[17,1061,1062],{},"Strategy",[129,1064,1065],{},"Consumer super app",[129,1067,449],{},[129,1069,452],{},[129,1071,455],{},[13,1073,1074,1075,1079],{},"OpenAI chooses to be the model of the consumer surface. Claude remains the model of the IDE. Gemini is the model of enterprise cloud, as confirmed by ",[87,1076,1078],{"href":1077},"\u002Fnews\u002Fgoogle-cloud-next-2026-gemini-enterprise-agent-platform-a2a-tpu8-april-2026","Google Cloud Next 2026 with the Gemini Enterprise Agent Platform and the A2A protocol in production at 150 organizations",". And Qwen is the model of teams that refuse to pay for an API. The landscape finally segments by use case, not by \"who has the biggest benchmark.\"",[46,1081,1083],{"id":1082},"what-it-means-for-developers-and-publishers","What it means for developers and publishers",[13,1085,1086],{},"For developers building AI apps, three consequences:",[13,1088,1089,1092],{},[17,1090,1091],{},"Prompt engineering for the agent becomes a skill."," GPT-5.5 excels at under-specified tasks but collapses on badly described orchestrations. The gap between an agent that delivers and an agent that loops widens.",[13,1094,1095,1098,1099,487],{},[17,1096,1097],{},"Per-token price stops being the relevant KPI."," Product teams will track \"price per task completed,\" which includes retries, context, memory. The teams that already figured this out are the ones ",[87,1100,1102],{"href":1101},"\u002Fpublishers","monetizing their AI apps with @idlen\u002Fchat-sdk — where every token the user doesn't use doesn't disappear, it capitalizes",[13,1104,1105,1108,1109,1113],{},[17,1106,1107],{},"The agent becomes the subscription unit."," ChatGPT Pro at $200\u002Fmonth with unlimited GPT-5.5 Pro isn't a price, it's a subscription to a permanent agent's production. ",[87,1110,1112],{"href":1111},"\u002Fnews\u002Fcursor-3-agent-first-interface-parallel-agents-vibe-coding-2026","Lovable, Cursor, and Emergent already follow this line",". OpenAI just normalized it on the consumer side.",[500,1115],{},[13,1117,1118],{},[54,1119,1120],{},[17,1121,1122],{},"In summary:",[59,1124,1125,1135,1141,1147,1153,1159],{},[62,1126,1127,1130,1131,1134],{},[17,1128,1129],{},"OpenAI shipped GPT-5.5"," on ",[17,1132,1133],{},"April 23, 2026"," — first fully retrained base model since GPT-4.5.",[62,1136,1137,1140],{},[17,1138,1139],{},"Key benchmarks",": 84.9% on GDPval, 78.7% on OSWorld-Verified, 82.7% on Terminal-Bench 2.0, 98.0% on Tau2-bench Telecom, 58.6% on SWE-Bench Pro.",[62,1142,1143,1146],{},[17,1144,1145],{},"API pricing doubled",": $5 \u002F $30 per million tokens (with the GPT-5.5 Pro variant at $30 \u002F $180).",[62,1148,1149,1152],{},[17,1150,1151],{},"The super app"," merges ChatGPT, Codex, and Atlas under a single model and a single session.",[62,1154,1155,1158],{},[17,1156,1157],{},"Available day one"," on Plus, Pro, Business, Enterprise, and the API.",[62,1160,1161,1164],{},[17,1162,1163],{},"Positioning",": OpenAI cedes SWE-Bench to Claude and claims pole position on the computer as the agent's playing field.",[500,1166],{},[13,1168,1169],{},"GPT-5.5 is a bet on form, not on the model. OpenAI has figured out that raw intelligence no longer wins market share — what matters is the integration into the daily life of the developer, the knowledge worker, the researcher. The ChatGPT + Codex + Atlas super app is the embodiment of that bet. Whether enterprises will pay twice the price for a model that scores 58% on SWE-Bench when Claude Opus 4.7 scores 87% is the question. The answer will come with Q2 2026 revenue numbers.",[13,1171,1172],{},"Sources:",[59,1174,1175,1180,1185,1190,1195,1200],{},[62,1176,1177],{},[87,1178,568],{"href":565,"rel":1179},[567],[62,1181,1182],{},[87,1183,575],{"href":573,"rel":1184},[567],[62,1186,1187],{},[87,1188,582],{"href":580,"rel":1189},[567],[62,1191,1192],{},[87,1193,589],{"href":587,"rel":1194},[567],[62,1196,1197],{},[87,1198,596],{"href":594,"rel":1199},[567],[62,1201,1202],{},[87,1203,603],{"href":601,"rel":1204},[567],{"title":605,"searchDepth":606,"depth":606,"links":1206},[1207,1208,1209,1210,1211,1212,1213],{"id":702,"depth":606,"text":703},{"id":745,"depth":606,"text":746},{"id":839,"depth":606,"text":840},{"id":916,"depth":606,"text":917},{"id":941,"depth":606,"text":942},{"id":960,"depth":606,"text":961},{"id":1082,"depth":606,"text":1083},"On April 23, 2026, OpenAI unveiled GPT-5.5, its first fully retrained base model since GPT-4.5. 84.9% on GDPval, 78.7% on OSWorld, 82.7% on Terminal-Bench 2.0 — with API pricing doubled to $5 \u002F $30 per million tokens. Goal: merge ChatGPT, Codex, and Atlas into a single agentic session.",{},"\u002Fnews\u002Fopenai-gpt-55-super-app-codex-atlas-agent-april-2026",{"title":663,"description":1214},"On April 23, 2026, OpenAI launches GPT-5.5 — 84.9% GDPval, 78.7% OSWorld, 82.7% Terminal-Bench 2.0. First fully retrained base model since GPT-4.5. API pricing at $5 \u002F $30 per million tokens. The ChatGPT + Codex + Atlas super app takes shape.","OpenAI GPT-5.5: Agentic Super App ChatGPT + Codex + Atlas — April 2026","news\u002Fopenai-gpt-55-super-app-codex-atlas-agent-april-2026",[628,629,630,631,632,633,634,635,636,637],"b1_YdM3Sy57Mq-MIRYCsu7cGuJFDJU2IoAKD1i1mKu8",{"id":5,"title":6,"author":7,"author_avatar":8,"body":1224,"category":615,"date":616,"description":617,"extension":618,"image":619,"meta":1637,"navigation":621,"path":622,"seo":1638,"seoDescription":624,"seoTitle":625,"stem":626,"tags":1639,"translationKey":638,"updatedAt":8,"__hash__":639},{"type":10,"value":1225,"toc":1628},[1226,1242,1244,1248,1264,1270,1272,1274,1340,1346,1348,1352,1404,1408,1410,1412,1416,1422,1426,1428,1432,1436,1438,1530,1534,1536,1538,1542,1548,1554,1556,1562,1590,1592,1594,1596],[13,1227,15,1228,20,1230,24,1232,28,1234,32,1236,36,1238,40,1240,44],{},[17,1229,19],{},[17,1231,23],{},[17,1233,27],{},[17,1235,31],{},[17,1237,35],{},[17,1239,39],{},[17,1241,43],{},[46,1243,49],{"id":48},[13,1245,52,1246,57],{},[54,1247,56],{},[59,1249,1250,1254,1260],{},[62,1251,1252,67],{},[17,1253,66],{},[62,1255,1256,73,1258],{},[17,1257,72],{},[54,1259,76],{},[62,1261,1262,82],{},[17,1263,81],{},[13,1265,85,1266,91,1268,96],{},[87,1267,90],{"href":89},[87,1269,95],{"href":94},[46,1271,100],{"id":99},[13,1273,103],{},[105,1275,1276,1286],{},[108,1277,1278],{},[111,1279,1280,1282,1284],{},[114,1281,116],{},[114,1283,119],{},[114,1285,122],{},[124,1287,1288,1298,1308,1318,1328],{},[111,1289,1290,1294,1296],{},[129,1291,1292],{},[17,1293,133],{},[129,1295,27],{},[129,1297,138],{},[111,1299,1300,1304,1306],{},[129,1301,1302],{},[17,1303,145],{},[129,1305,31],{},[129,1307,150],{},[111,1309,1310,1314,1316],{},[129,1311,1312],{},[17,1313,157],{},[129,1315,35],{},[129,1317,162],{},[111,1319,1320,1324,1326],{},[129,1321,1322],{},[17,1323,169],{},[129,1325,172],{},[129,1327,175],{},[111,1329,1330,1334,1336],{},[129,1331,1332],{},[17,1333,182],{},[129,1335,39],{},[129,1337,187,1338],{},[54,1339,190],{},[13,1341,193,1342,198,1344,203],{},[87,1343,197],{"href":196},[87,1345,202],{"href":201},[46,1347,207],{"id":206},[13,1349,210,1350,214],{},[17,1351,213],{},[105,1353,1354,1364],{},[108,1355,1356],{},[111,1357,1358,1360,1362],{},[114,1359,223],{},[114,1361,226],{},[114,1363,229],{},[124,1365,1366,1376,1390],{},[111,1367,1368,1372,1374],{},[129,1369,1370],{},[17,1371,238],{},[129,1373,241],{},[129,1375,244],{},[111,1377,1378,1382,1386],{},[129,1379,1380],{},[17,1381,19],{},[129,1383,1384],{},[17,1385,255],{},[129,1387,1388],{},[17,1389,260],{},[111,1391,1392,1396,1400],{},[129,1393,1394],{},[17,1395,267],{},[129,1397,1398],{},[17,1399,260],{},[129,1401,1402],{},[17,1403,276],{},[13,1405,279,1406,284],{},[87,1407,283],{"href":282},[46,1409,288],{"id":287},[13,1411,291],{},[13,1413,1414,297],{},[17,1415,296],{},[13,1417,1418,303,1420,307],{},[17,1419,302],{},[54,1421,306],{},[13,1423,1424,313],{},[17,1425,312],{},[46,1427,317],{"id":316},[13,1429,320,1430,324],{},[17,1431,323],{},[13,1433,327,1434,332],{},[87,1435,331],{"href":330},[46,1437,336],{"id":335},[105,1439,1440,1454],{},[108,1441,1442],{},[111,1443,1444,1446,1448,1450,1452],{},[114,1445,345],{},[114,1447,19],{},[114,1449,350],{},[114,1451,353],{},[114,1453,356],{},[124,1455,1456,1472,1488,1502,1516],{},[111,1457,1458,1462,1464,1468,1470],{},[129,1459,1460],{},[17,1461,365],{},[129,1463,368],{},[129,1465,1466],{},[17,1467,373],{},[129,1469,376],{},[129,1471,379],{},[111,1473,1474,1478,1482,1484,1486],{},[129,1475,1476],{},[17,1477,386],{},[129,1479,1480],{},[17,1481,391],{},[129,1483,394],{},[129,1485,397],{},[129,1487,400],{},[111,1489,1490,1494,1496,1498,1500],{},[129,1491,1492],{},[17,1493,407],{},[129,1495,410],{},[129,1497,413],{},[129,1499,416],{},[129,1501,419],{},[111,1503,1504,1508,1510,1512,1514],{},[129,1505,1506],{},[17,1507,426],{},[129,1509,429],{},[129,1511,429],{},[129,1513,429],{},[129,1515,436],{},[111,1517,1518,1522,1524,1526,1528],{},[129,1519,1520],{},[17,1521,443],{},[129,1523,446],{},[129,1525,449],{},[129,1527,452],{},[129,1529,455],{},[13,1531,458,1532,463],{},[87,1533,462],{"href":461},[46,1535,467],{"id":466},[13,1537,470],{},[13,1539,1540,476],{},[17,1541,475],{},[13,1543,1544,482,1546,487],{},[17,1545,481],{},[87,1547,486],{"href":485},[13,1549,1550,493,1552,498],{},[17,1551,492],{},[87,1553,497],{"href":496},[500,1555],{},[13,1557,1558],{},[54,1559,1560],{},[17,1561,508],{},[59,1563,1564,1570,1574,1578,1582,1586],{},[62,1565,1566,516,1568,520],{},[17,1567,515],{},[17,1569,519],{},[62,1571,1572,526],{},[17,1573,525],{},[62,1575,1576,532],{},[17,1577,531],{},[62,1579,1580,538],{},[17,1581,537],{},[62,1583,1584,544],{},[17,1585,543],{},[62,1587,1588,550],{},[17,1589,549],{},[500,1591],{},[13,1593,555],{},[13,1595,558],{},[59,1597,1598,1603,1608,1613,1618,1623],{},[62,1599,1600],{},[87,1601,568],{"href":565,"rel":1602},[567],[62,1604,1605],{},[87,1606,575],{"href":573,"rel":1607},[567],[62,1609,1610],{},[87,1611,582],{"href":580,"rel":1612},[567],[62,1614,1615],{},[87,1616,589],{"href":587,"rel":1617},[567],[62,1619,1620],{},[87,1621,596],{"href":594,"rel":1622},[567],[62,1624,1625],{},[87,1626,603],{"href":601,"rel":1627},[567],{"title":605,"searchDepth":606,"depth":606,"links":1629},[1630,1631,1632,1633,1634,1635,1636],{"id":48,"depth":606,"text":49},{"id":99,"depth":606,"text":100},{"id":206,"depth":606,"text":207},{"id":287,"depth":606,"text":288},{"id":316,"depth":606,"text":317},{"id":335,"depth":606,"text":336},{"id":466,"depth":606,"text":467},{},{"title":6,"description":617},[628,629,630,631,632,633,634,635,636,637]]