[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-post-\u002Ffr\u002Fnews\u002Fgpt-54-controle-ordinateur-openai-agent-autonome":3,"news-related-fr-\u002Ffr\u002Fnews\u002Fgpt-54-controle-ordinateur-openai-agent-autonome":437,"hreflang-gpt-54-computer-control":457},{"doc":4,"redirect":8},{"id":5,"title":6,"author":7,"author_avatar":8,"body":9,"category":414,"date":415,"description":416,"extension":417,"image":418,"meta":419,"navigation":420,"path":421,"seo":422,"seoDescription":423,"seoTitle":424,"stem":425,"tags":426,"translationKey":435,"updatedAt":8,"__hash__":436},"news_fr\u002Ffr\u002Fnews\u002Fgpt-54-controle-ordinateur-openai-agent-autonome.md","GPT-5.4 peut contrôler votre ordinateur : OpenAI franchit une nouvelle frontière","Paul Lefizelier",null,{"type":10,"value":11,"toc":404},"minimark",[12,41,46,61,76,86,90,100,103,117,125,129,132,144,149,260,270,274,281,304,311,319,323,334,337,344,356,359,363,396,398],[13,14,15,16,20,21,24,25,28,29,32,33,36,37,40],"p",{},"Ce 16 mars 2026, ",[17,18,19],"strong",{},"OpenAI"," lance ",[17,22,23],{},"GPT-5.4",", un modèle capable de ",[17,26,27],{},"contrôler un ordinateur de façon autonome",". Ouvrir des applications, cliquer, remplir des formulaires, naviguer sur le web — le tout sans intervention humaine. GPT-5.4 établit de nouveaux records sur les benchmarks ",[17,30,31],{},"OSWorld-Verified"," et ",[17,34,35],{},"WebArena",", avec ",[17,38,39],{},"18 % d'erreurs en moins"," que GPT-5.2. Trois versions sont disponibles : standard, Thinking et Pro.",[42,43,45],"h2",{"id":44},"gpt-54-bien-plus-quun-chatbot","GPT-5.4 : bien plus qu'un chatbot",[13,47,48,49,56,57,60],{},"GPT-5.4 n'est pas une simple mise à jour incrémentale. C'est le premier modèle d'",[50,51,19],"a",{"href":52,"rel":53,"target":55},"https:\u002F\u002Fopenai.com",[54],"nofollow","\\_blank"," conçu dès le départ pour le ",[17,58,59],{},"computer use"," — la capacité d'un agent IA à contrôler un ordinateur comme le ferait un humain.",[13,62,63,64,67,68,71,72,75],{},"Le modèle est disponible en trois versions. ",[17,65,66],{},"GPT-5.4 standard"," couvre les tâches courantes. ",[17,69,70],{},"GPT-5.4 Thinking"," ajoute un raisonnement approfondi pour les tâches complexes. ",[17,73,74],{},"GPT-5.4 Pro"," est la version la plus performante, destinée aux workflows professionnels exigeants.",[13,77,78,79,82,83,85],{},"Les trois versions partagent une fenêtre de contexte d'",[17,80,81],{},"1 million de tokens",". Par rapport à GPT-5.2, le gain principal est la fiabilité : ",[17,84,39],{}," sur les tâches d'interaction avec un desktop. Ce n'est plus un prototype. C'est un outil utilisable en production.",[42,87,89],{"id":88},"comment-fonctionne-le-computer-use-concrètement","Comment fonctionne le computer use concrètement ?",[13,91,92,93,95,96,99],{},"Le ",[17,94,59],{}," (contrôle autonome d'ordinateur) repose sur un principe simple. L'agent capture des ",[17,97,98],{},"screenshots"," de l'écran, identifie les éléments cliquables — boutons, champs de texte, menus — puis exécute des actions : clic, frappe, scroll, navigation entre applications.",[13,101,102],{},"Concrètement, vous pouvez lui donner une instruction comme : « Réserve un vol Paris-New York pour le 15 avril, envoie la confirmation par email à Thomas et ajoute le voyage au calendrier. » GPT-5.4 exécute chaque étape en séquence, sans que vous touchiez le clavier.",[13,104,105,106,109,110,32,113,116],{},"Le modèle fonctionne sur ",[17,107,108],{},"macOS",", ",[17,111,112],{},"Windows",[17,114,115],{},"Linux",". Il s'adapte à l'interface de chaque système d'exploitation et reconnaît les éléments visuels même si leur position change.",[13,118,119,120,124],{},"Pour les développeurs, le cas d'usage le plus parlant : un agent qui ouvre ",[50,121,123],{"href":122},"\u002Ffr\u002Fnews\u002Fvibecoding-outils-agentic-ai-dev-2026","Cursor ou un IDE IA",", écrit du code, lance un build, lit les erreurs de compilation et les corrige — en boucle, sans intervention.",[42,126,128],{"id":127},"les-benchmarks-gpt-54-dépasse-les-humains-sur-les-tâches-desktop","Les benchmarks : GPT-5.4 dépasse les humains sur les tâches desktop",[13,130,131],{},"GPT-5.4 établit de nouveaux records sur les deux benchmarks de référence pour le computer use.",[13,133,134,135,139,140,143],{},"Sur ",[50,136,31],{"href":137,"rel":138,"target":55},"https:\u002F\u002Fos-world.github.io",[54],", le benchmark qui évalue la capacité d'un agent à réaliser des tâches sur un vrai système d'exploitation, GPT-5.4 dépasse le score humain de référence fixé à ",[17,141,142],{},"72,4 %",". C'est la première fois qu'un modèle franchit ce seuil.",[13,145,134,146,148],{},[17,147,35],{},", qui mesure la performance sur des tâches web autonomes (navigation, remplissage de formulaires, interactions multi-sites), GPT-5.4 établit également un nouveau record.",[150,151,152,174],"table",{},[153,154,155],"thead",{},[156,157,158,162,165,168,171],"tr",{},[159,160,161],"th",{},"Solution",[159,163,164],{},"Éditeur",[159,166,167],{},"Lancement",[159,169,170],{},"Précision",[159,172,173],{},"Dispo",[175,176,177,195,212,228,245],"tbody",{},[156,178,179,184,186,189,192],{},[180,181,182],"td",{},[17,183,23],{},[180,185,19],{},[180,187,188],{},"Mars 2026",[180,190,191],{},"Record OSWorld",[180,193,194],{},"API + Pro",[156,196,197,200,203,206,209],{},[180,198,199],{},"Claude computer use",[180,201,202],{},"Anthropic",[180,204,205],{},"Oct 2024",[180,207,208],{},"Bonne",[180,210,211],{},"API",[156,213,214,217,220,223,225],{},[180,215,216],{},"Project Mariner",[180,218,219],{},"Google",[180,221,222],{},"Déc 2024",[180,224,208],{},[180,226,227],{},"Beta",[156,229,230,233,236,239,242],{},[180,231,232],{},"Copilot Actions",[180,234,235],{},"Microsoft",[180,237,238],{},"Jan 2026",[180,240,241],{},"Correcte",[180,243,244],{},"M365",[156,246,247,250,252,255,258],{},[180,248,249],{},"Operator",[180,251,19],{},[180,253,254],{},"Jan 2025",[180,256,257],{},"Précédente",[180,259,211],{},[13,261,262,263,265,266,269],{},"Le précédent notable : ",[17,264,202],{}," avait lancé le computer use en beta en octobre 2024 avec ",[17,267,268],{},"Claude 3.5 Sonnet",". GPT-5.4 va significativement plus loin en précision et en fiabilité sur les tâches multi-étapes.",[42,271,273],{"id":272},"la-guerre-du-computer-use-openai-anthropic-google-microsoft","La guerre du computer use : OpenAI, Anthropic, Google, Microsoft",[13,275,276,277,280],{},"Le computer use est devenu la ",[17,278,279],{},"prochaine grande bataille IA"," après les LLM. Tous les géants s'y positionnent.",[13,282,283,285,286,288,289,291,292,294,295,297,298,300,301,303],{},[17,284,202],{}," a été le premier à lancer une beta publique avec Claude computer use. ",[17,287,219],{}," a suivi avec ",[17,290,216],{},", un agent de navigation web intégré à Chrome. ",[17,293,235],{}," pousse ",[17,296,232],{}," dans Microsoft 365, ciblant les tâches bureautiques. Et ",[17,299,19],{}," avait déjà lancé ",[17,302,249],{}," en janvier 2025, une première itération plus limitée.",[13,305,306,307,310],{},"Avec GPT-5.4, ",[17,308,309],{},"Sam Altman"," et OpenAI prennent la tête de la course. Le record sur OSWorld est un signal fort : GPT-5.4 n'est pas seulement le meilleur chatbot. C'est l'agent autonome le plus fiable du marché.",[13,312,313,314,318],{},"L'impact touche directement les métiers. Assistants virtuels, support client, ops, finance, RH — toutes les tâches répétitives sur ordinateur deviennent automatisables. Les ",[50,315,317],{"href":316},"\u002Ffr\u002Fblog\u002Fagents-ia-developpeurs-guide-outils-autonomes-2026","agents IA autonomes"," passent du concept au produit.",[42,320,322],{"id":321},"ce-que-ça-change-pour-les-devs-et-builders","Ce que ça change pour les devs et builders",[13,324,325,326,329,330,333],{},"Pour les développeurs, GPT-5.4 ouvre un scénario longtemps théorique : le ",[17,327,328],{},"workflow de développement quasi autonome",". Un agent qui contrôle un ordinateur peut faire tourner ",[50,331,332],{"href":122},"Cursor ou Replit",", lancer des tests, déployer du code, gérer des outils SaaS — sans humain dans la boucle.",[13,335,336],{},"Combiné à un système de mémoire persistante comme Nyne, un agent GPT-5.4 qui connaît votre stack, vos conventions et vos préférences devient un collaborateur redoutable.",[13,338,339,340,343],{},"Les limites existent. Le taux d'erreur n'est pas nul — 18 % de mieux que GPT-5.2, mais pas infaillible. La ",[17,341,342],{},"sécurité"," pose question : un contenu affiché à l'écran peut potentiellement injecter des instructions malveillantes dans l'agent (prompt injection visuel). L'accès aux apps nécessitant une authentification biométrique reste impossible. Et le tarif Pro n'est pas accessible au free tier.",[13,345,346,347,351,352,355],{},"Malgré ces réserves, la direction est claire. Les ",[50,348,350],{"href":349},"\u002Ffr\u002Fnews\u002Fdeepseek-v4-modele-ia-open-weight-1000-milliards","modèles frontier comme DeepSeek V4"," se battent sur le raisonnement. GPT-5.4 se bat sur l'",[17,353,354],{},"action",".",[357,358],"hr",{},[42,360,362],{"id":361},"en-résumé","En résumé",[364,365,366,372,378,384,390],"ul",{},[367,368,369,371],"li",{},[17,370,23],{}," est le nouveau modèle d'OpenAI capable de contrôler un ordinateur de façon autonome : clics, formulaires, navigation web et tâches multi-étapes.",[367,373,374,377],{},[17,375,376],{},"Il établit des records"," sur les benchmarks OSWorld-Verified (dépassant le score humain de 72,4 %) et WebArena, avec 18 % d'erreurs en moins que GPT-5.2.",[367,379,380,383],{},[17,381,382],{},"Trois versions"," sont disponibles : GPT-5.4 standard, Thinking et Pro, avec une fenêtre de contexte d'1 million de tokens.",[367,385,386,389],{},[17,387,388],{},"La bataille du computer use"," oppose OpenAI, Anthropic (Claude), Google (Project Mariner) et Microsoft (Copilot Actions) — GPT-5.4 prend la tête.",[367,391,392,395],{},[17,393,394],{},"Les limites"," incluent un taux d'erreur résiduel, des risques de prompt injection visuel et un tarif Pro élevé non accessible gratuitement.",[357,397],{},[13,399,400],{},[401,402,403],"em",{},"Si un agent peut contrôler votre ordinateur mieux que vous sur les tâches répétitives — réserver, remplir, déployer, tester — quel est encore le rôle irremplaçable de l'humain dans un workflow de développement ? Probablement le même qu'il a toujours été : décider quoi construire, et pourquoi.",{"title":405,"searchDepth":406,"depth":406,"links":407},"",2,[408,409,410,411,412,413],{"id":44,"depth":406,"text":45},{"id":88,"depth":406,"text":89},{"id":127,"depth":406,"text":128},{"id":272,"depth":406,"text":273},{"id":321,"depth":406,"text":322},{"id":361,"depth":406,"text":362},"ai","2026-03-16T16:19:00","OpenAI lance GPT-5.4, un agent IA capable de contrôler un ordinateur de façon autonome. Record sur OSWorld, 3 versions et computer use avancé.","md","\u002Fgpt-54-controle-ordinateur-openai-agent-autonome.jpg",{},true,"\u002Ffr\u002Fnews\u002Fgpt-54-controle-ordinateur-openai-agent-autonome",{"title":6,"description":416},"GPT-5.4 d'OpenAI contrôle un ordinateur de façon autonome : clics, formulaires, navigation web. Record OSWorld, 18 % d'erreurs en moins que GPT-5.2.","GPT-5.4 : OpenAI lance le contrôle autonome d'ordinateur","fr\u002Fnews\u002Fgpt-54-controle-ordinateur-openai-agent-autonome",[427,428,429,430,431,432,433,434],"gpt-54","openai","computer-use","agent-autonome","agentic-ai","osworld","webArena","llm","gpt-54-computer-control","zZa2uFEM96GnxAmskNelWqs4jsQ73TtmYPX45XgMiO4",[438,445,451],{"title":439,"description":440,"path":441,"date":442,"image":443,"category":444},"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":446,"description":447,"path":448,"date":449,"image":450,"category":414},"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":452,"description":453,"path":454,"date":455,"image":456,"category":414},"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":458,"fr":825,"es":8},{"id":459,"title":460,"author":7,"author_avatar":8,"body":461,"category":414,"date":415,"description":815,"extension":417,"image":418,"meta":816,"navigation":420,"path":817,"seo":818,"seoDescription":819,"seoTitle":820,"stem":821,"tags":822,"translationKey":435,"updatedAt":8,"__hash__":824},"news_en\u002Fnews\u002Fgpt-54-computer-control-openai-autonomous-agent.md","GPT-5.4 Can Control Your Computer: OpenAI Crosses a New Frontier",{"type":10,"value":462,"toc":807},[463,486,490,500,512,522,526,535,538,548,556,560,563,574,579,669,678,682,689,712,718,725,729,740,747,754,763,765,769,800,802],[13,464,465,466,468,469,471,472,475,476,478,479,481,482,485],{},"On March 16, 2026, ",[17,467,19],{}," launches ",[17,470,23],{},", a model capable of ",[17,473,474],{},"autonomously controlling a computer",". Opening apps, clicking, filling out forms, browsing the web — all without human intervention. GPT-5.4 sets new records on the ",[17,477,31],{}," and ",[17,480,35],{}," benchmarks, with ",[17,483,484],{},"18% fewer errors"," than GPT-5.2. Three versions are available: standard, Thinking and Pro.",[42,487,489],{"id":488},"gpt-54-far-more-than-a-chatbot","GPT-5.4: Far More Than a Chatbot",[13,491,492,493,496,497,499],{},"GPT-5.4 is not a simple incremental update. It's the first ",[50,494,19],{"href":52,"rel":495,"target":55},[54]," model designed from the ground up for ",[17,498,59],{}," — the ability for an AI agent to control a computer just like a human would.",[13,501,502,503,505,506,508,509,511],{},"The model comes in three versions. ",[17,504,66],{}," handles common tasks. ",[17,507,70],{}," adds deep reasoning for complex workflows. ",[17,510,74],{}," is the most capable version, built for demanding professional use cases.",[13,513,514,515,518,519,521],{},"All three versions share a ",[17,516,517],{},"1-million token"," context window. Compared to GPT-5.2, the main improvement is reliability: ",[17,520,484],{}," on desktop interaction tasks. This is no longer a prototype. It's a production-ready tool.",[42,523,525],{"id":524},"how-does-computer-use-actually-work","How Does Computer Use Actually Work?",[13,527,528,531,532,534],{},[17,529,530],{},"Computer use"," (autonomous computer control) works on a simple principle. The agent captures ",[17,533,98],{}," of the screen, identifies clickable elements — buttons, text fields, menus — then executes actions: clicks, keystrokes, scrolling, switching between apps.",[13,536,537],{},"In practice, you can give it an instruction like: \"Book a Paris-to-New York flight for April 15, email the confirmation to Thomas and add the trip to the calendar.\" GPT-5.4 executes each step in sequence, without you touching the keyboard.",[13,539,540,541,109,543,478,545,547],{},"The model works on ",[17,542,108],{},[17,544,112],{},[17,546,115],{},". It adapts to each operating system's interface and recognizes visual elements even if their position changes.",[13,549,550,551,555],{},"For developers, the most compelling use case: an agent that opens ",[552,553,554],"span",{},"INTERNAL LINK: Cursor or an AI IDE | vibe coding article",", writes code, runs a build, reads compilation errors and fixes them — in a loop, without intervention.",[42,557,559],{"id":558},"the-benchmarks-gpt-54-surpasses-humans-on-desktop-tasks","The Benchmarks: GPT-5.4 Surpasses Humans on Desktop Tasks",[13,561,562],{},"GPT-5.4 sets new records on both reference benchmarks for computer use.",[13,564,565,566,569,570,573],{},"On ",[50,567,31],{"href":137,"rel":568,"target":55},[54],", the benchmark evaluating an agent's ability to complete tasks on a real operating system, GPT-5.4 exceeds the human reference score of ",[17,571,572],{},"72.4%",". It's the first time a model has crossed this threshold.",[13,575,565,576,578],{},[17,577,35],{},", which measures performance on autonomous web tasks (browsing, form filling, multi-site interactions), GPT-5.4 also sets a new record.",[150,580,581,599],{},[153,582,583],{},[156,584,585,587,590,593,596],{},[159,586,161],{},[159,588,589],{},"Publisher",[159,591,592],{},"Launch",[159,594,595],{},"Accuracy",[159,597,598],{},"Availability",[175,600,601,617,630,643,656],{},[156,602,603,607,609,612,615],{},[180,604,605],{},[17,606,23],{},[180,608,19],{},[180,610,611],{},"March 2026",[180,613,614],{},"OSWorld record",[180,616,194],{},[156,618,619,621,623,625,628],{},[180,620,199],{},[180,622,202],{},[180,624,205],{},[180,626,627],{},"Good",[180,629,211],{},[156,631,632,634,636,639,641],{},[180,633,216],{},[180,635,219],{},[180,637,638],{},"Dec 2024",[180,640,627],{},[180,642,227],{},[156,644,645,647,649,651,654],{},[180,646,232],{},[180,648,235],{},[180,650,238],{},[180,652,653],{},"Decent",[180,655,244],{},[156,657,658,660,662,664,667],{},[180,659,249],{},[180,661,19],{},[180,663,254],{},[180,665,666],{},"Previous gen",[180,668,211],{},[13,670,671,672,674,675,677],{},"The key precedent: ",[17,673,202],{}," launched computer use in beta in October 2024 with ",[17,676,268],{},". GPT-5.4 goes significantly further in accuracy and reliability on multi-step tasks.",[42,679,681],{"id":680},"the-computer-use-war-openai-anthropic-google-microsoft","The Computer Use War: OpenAI, Anthropic, Google, Microsoft",[13,683,684,685,688],{},"Computer use has become the ",[17,686,687],{},"next major AI battle"," after LLMs. Every major player is positioning for it.",[13,690,691,693,694,696,697,699,700,702,703,705,706,708,709,711],{},[17,692,202],{}," was first to launch a public beta with Claude computer use. ",[17,695,219],{}," followed with ",[17,698,216],{},", a web navigation agent integrated into Chrome. ",[17,701,235],{}," pushes ",[17,704,232],{}," into Microsoft 365, targeting office tasks. And ",[17,707,19],{}," had already launched ",[17,710,249],{}," in January 2025, a more limited first iteration.",[13,713,714,715,717],{},"With GPT-5.4, ",[17,716,309],{}," and OpenAI take the lead. The OSWorld record sends a strong signal: GPT-5.4 isn't just the best chatbot. It's the most reliable autonomous agent on the market.",[13,719,720,721,724],{},"The impact hits roles directly. Virtual assistants, customer support, ops, finance, HR — all repetitive computer tasks are now automatable. ",[552,722,723],{},"INTERNAL LINK: Autonomous AI agents | AI agents in enterprise article"," are moving from concept to product.",[42,726,728],{"id":727},"what-this-changes-for-devs-and-builders","What This Changes for Devs and Builders",[13,730,731,732,735,736,739],{},"For developers, GPT-5.4 opens a long-theorized scenario: the ",[17,733,734],{},"near-autonomous development workflow",". An agent that controls a computer can run ",[552,737,738],{},"INTERNAL LINK: Cursor or Replit | vibe coding article on Idlen",", launch tests, deploy code, manage SaaS tools — with no human in the loop.",[13,741,742,743,746],{},"Combined with a ",[552,744,745],{},"INTERNAL LINK: persistent memory system like Nyne | Nyne AI agent article",", a GPT-5.4 agent that knows your stack, conventions and preferences becomes a formidable collaborator.",[13,748,749,750,753],{},"Limitations exist. The error rate isn't zero — 18% better than GPT-5.2, but not flawless. ",[17,751,752],{},"Security"," raises concerns: content displayed on screen can potentially inject malicious instructions into the agent (visual prompt injection). Apps requiring biometric authentication remain inaccessible. And Pro pricing isn't available on the free tier.",[13,755,756,757,760,761,355],{},"Despite these caveats, the direction is clear. ",[552,758,759],{},"INTERNAL LINK: Frontier models like DeepSeek V4 | DeepSeek V4 article"," compete on reasoning. GPT-5.4 competes on ",[17,762,354],{},[357,764],{},[42,766,768],{"id":767},"key-takeaways","Key Takeaways",[364,770,771,776,782,788,794],{},[367,772,773,775],{},[17,774,23],{}," is OpenAI's new model capable of autonomously controlling a computer: clicks, forms, web browsing and multi-step tasks.",[367,777,778,781],{},[17,779,780],{},"It sets records"," on OSWorld-Verified (surpassing the 72.4% human score) and WebArena benchmarks, with 18% fewer errors than GPT-5.2.",[367,783,784,787],{},[17,785,786],{},"Three versions"," are available: GPT-5.4 standard, Thinking and Pro, with a 1-million token context window.",[367,789,790,793],{},[17,791,792],{},"The computer use battle"," pits OpenAI, Anthropic (Claude), Google (Project Mariner) and Microsoft (Copilot Actions) — GPT-5.4 takes the lead.",[367,795,796,799],{},[17,797,798],{},"Limitations"," include a residual error rate, visual prompt injection risks and high Pro pricing not available on the free tier.",[357,801],{},[13,803,804],{},[401,805,806],{},"If an agent can control your computer better than you on repetitive tasks — booking, filling, deploying, testing — what is the irreplaceable role of the human in a development workflow? Probably the same it's always been: deciding what to build, and why.",{"title":405,"searchDepth":406,"depth":406,"links":808},[809,810,811,812,813,814],{"id":488,"depth":406,"text":489},{"id":524,"depth":406,"text":525},{"id":558,"depth":406,"text":559},{"id":680,"depth":406,"text":681},{"id":727,"depth":406,"text":728},{"id":767,"depth":406,"text":768},"OpenAI launches GPT-5.4, an AI agent that autonomously controls a computer. Record OSWorld scores, 3 versions and advanced computer use.",{},"\u002Fnews\u002Fgpt-54-computer-control-openai-autonomous-agent",{"title":460,"description":815},"OpenAI's GPT-5.4 autonomously controls a computer: clicks, forms, web browsing. Record OSWorld score, 18% fewer errors than GPT-5.2. Full breakdown.","GPT-5.4: OpenAI launches autonomous computer control","news\u002Fgpt-54-computer-control-openai-autonomous-agent",[427,428,429,823,431,432,433,434],"autonomous-agent","BK1skA7d1dLqei8j7F8DW45buyi6Pzzfr9Qjg_GyfdA",{"id":5,"title":6,"author":7,"author_avatar":8,"body":826,"category":414,"date":415,"description":416,"extension":417,"image":418,"meta":1078,"navigation":420,"path":421,"seo":1079,"seoDescription":423,"seoTitle":424,"stem":425,"tags":1080,"translationKey":435,"updatedAt":8,"__hash__":436},{"type":10,"value":827,"toc":1070},[828,842,844,851,859,865,867,873,875,883,887,889,891,898,902,982,988,990,994,1010,1014,1018,1020,1026,1028,1032,1038,1040,1042,1064,1066],[13,829,15,830,20,832,24,834,28,836,32,838,36,840,40],{},[17,831,19],{},[17,833,23],{},[17,835,27],{},[17,837,31],{},[17,839,35],{},[17,841,39],{},[42,843,45],{"id":44},[13,845,48,846,56,849,60],{},[50,847,19],{"href":52,"rel":848,"target":55},[54],[17,850,59],{},[13,852,63,853,67,855,71,857,75],{},[17,854,66],{},[17,856,70],{},[17,858,74],{},[13,860,78,861,82,863,85],{},[17,862,81],{},[17,864,39],{},[42,866,89],{"id":88},[13,868,92,869,95,871,99],{},[17,870,59],{},[17,872,98],{},[13,874,102],{},[13,876,105,877,109,879,32,881,116],{},[17,878,108],{},[17,880,112],{},[17,882,115],{},[13,884,119,885,124],{},[50,886,123],{"href":122},[42,888,128],{"id":127},[13,890,131],{},[13,892,134,893,139,896,143],{},[50,894,31],{"href":137,"rel":895,"target":55},[54],[17,897,142],{},[13,899,134,900,148],{},[17,901,35],{},[150,903,904,918],{},[153,905,906],{},[156,907,908,910,912,914,916],{},[159,909,161],{},[159,911,164],{},[159,913,167],{},[159,915,170],{},[159,917,173],{},[175,919,920,934,946,958,970],{},[156,921,922,926,928,930,932],{},[180,923,924],{},[17,925,23],{},[180,927,19],{},[180,929,188],{},[180,931,191],{},[180,933,194],{},[156,935,936,938,940,942,944],{},[180,937,199],{},[180,939,202],{},[180,941,205],{},[180,943,208],{},[180,945,211],{},[156,947,948,950,952,954,956],{},[180,949,216],{},[180,951,219],{},[180,953,222],{},[180,955,208],{},[180,957,227],{},[156,959,960,962,964,966,968],{},[180,961,232],{},[180,963,235],{},[180,965,238],{},[180,967,241],{},[180,969,244],{},[156,971,972,974,976,978,980],{},[180,973,249],{},[180,975,19],{},[180,977,254],{},[180,979,257],{},[180,981,211],{},[13,983,262,984,265,986,269],{},[17,985,202],{},[17,987,268],{},[42,989,273],{"id":272},[13,991,276,992,280],{},[17,993,279],{},[13,995,996,285,998,288,1000,291,1002,294,1004,297,1006,300,1008,303],{},[17,997,202],{},[17,999,219],{},[17,1001,216],{},[17,1003,235],{},[17,1005,232],{},[17,1007,19],{},[17,1009,249],{},[13,1011,306,1012,310],{},[17,1013,309],{},[13,1015,313,1016,318],{},[50,1017,317],{"href":316},[42,1019,322],{"id":321},[13,1021,325,1022,329,1024,333],{},[17,1023,328],{},[50,1025,332],{"href":122},[13,1027,336],{},[13,1029,339,1030,343],{},[17,1031,342],{},[13,1033,346,1034,351,1036,355],{},[50,1035,350],{"href":349},[17,1037,354],{},[357,1039],{},[42,1041,362],{"id":361},[364,1043,1044,1048,1052,1056,1060],{},[367,1045,1046,371],{},[17,1047,23],{},[367,1049,1050,377],{},[17,1051,376],{},[367,1053,1054,383],{},[17,1055,382],{},[367,1057,1058,389],{},[17,1059,388],{},[367,1061,1062,395],{},[17,1063,394],{},[357,1065],{},[13,1067,1068],{},[401,1069,403],{},{"title":405,"searchDepth":406,"depth":406,"links":1071},[1072,1073,1074,1075,1076,1077],{"id":44,"depth":406,"text":45},{"id":88,"depth":406,"text":89},{"id":127,"depth":406,"text":128},{"id":272,"depth":406,"text":273},{"id":321,"depth":406,"text":322},{"id":361,"depth":406,"text":362},{},{"title":6,"description":416},[427,428,429,430,431,432,433,434]]