{"source":"Quelle IA, le classement des modèles de langage","page":"https://quelleia.com/benchmarks/metr_time_horizons/","edition":"2026-W40","date":"2026-09-28","echelle":"score_publie est le score de la source, dans son unité. score_0_1 est ce score ramené entre 0 et 1 pour le calcul. niveau_suggere est le score IA que cette seule mesure indique : 100 pour le meilleur modèle au lancement du site, 50 pour GPT-4o de novembre 2024. difficulte est le score IA d'un modèle qui obtient la moitié (score_0_1 = 0,5), pente la raideur de la courbe.","licence":"Réutilisation libre en citant « Quelle IA, quelleia.com » et la date de l'édition. Les mesures renvoient à leurs sources d'origine.","benchmark":{"id":"metr_time_horizons","nom":"METR Time Horizons (taux de réussite)","tache":"agents","etat":"actif","echelle":"unite","modeles":41,"poids_general":1.67,"difficulte":62,"pente":0.0448,"en_une_phrase":"Des tâches de logiciel, d'apprentissage automatique et de cybersécurité, dont on connaît la durée pour un professionnel humain.","notation":"Score moyen de réussite, de 0 à 1, sur l'ensemble des tâches de METR ; c'est la valeur qu'Epoch retient pour son indice.","limites":"Ce score moyen est moins parlant que l'horizon en minutes publié par METR. Les tâches sont propres et bien définies, loin du désordre d'un vrai poste.","mainteneur":"METR","url":"https://metr.org/time-horizons/","verifie":true},"tache":{"id":"agents","nom":"Agents","poids_general":15,"benchmarks_actifs":9},"modeles_mesures":41,"mesures":50,"classement":[{"rang":1,"modele":"claude-mythos-preview","nom":"Claude Mythos Preview","editeur":"Anthropic","page":"https://quelleia.com/modeles/claude-mythos-preview/","score_publie":0.85205,"score_0_1":0.852,"niveau_suggere":101.1,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/"},{"rang":2,"modele":"claude-opus-4-6","nom":"Claude Opus 4.6","editeur":"Anthropic","page":"https://quelleia.com/modeles/claude-opus-4-6/","score_publie":0.788644,"score_0_1":0.7886,"niveau_suggere":91.4,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":3,"modele":"gemini-3-1-pro","nom":"Gemini 3.1 Pro","editeur":"Google DeepMind","page":"https://quelleia.com/modeles/gemini-3-1-pro/","score_publie":0.77035,"score_0_1":0.7703,"niveau_suggere":89.1,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/"},{"rang":4,"modele":"gpt-5-2","nom":"GPT-5.2","editeur":"OpenAI","page":"https://quelleia.com/modeles/gpt-5-2/","score_publie":0.752875,"score_0_1":0.7529,"niveau_suggere":86.9,"reflexion":"elevee","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":5,"modele":"claude-opus-4-5","nom":"Claude Opus 4.5","editeur":"Anthropic","page":"https://quelleia.com/modeles/claude-opus-4-5/","score_publie":0.74969,"score_0_1":0.7497,"niveau_suggere":86.5,"reflexion":"budget","harnais":null,"configurations_mesurees":2,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":6,"modele":"gpt-5-3-codex","nom":"GPT-5.3 Codex","editeur":"OpenAI","page":"https://quelleia.com/modeles/gpt-5-3-codex/","score_publie":0.745439,"score_0_1":0.7454,"niveau_suggere":86,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":7,"modele":"gpt-5-4","nom":"GPT-5.4","editeur":"OpenAI","page":"https://quelleia.com/modeles/gpt-5-4/","score_publie":0.743443,"score_0_1":0.7434,"niveau_suggere":85.8,"reflexion":"inconnue","harnais":null,"configurations_mesurees":2,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":8,"modele":"gemini-3-pro","nom":"Gemini 3 Pro","editeur":"Google DeepMind","page":"https://quelleia.com/modeles/gemini-3-pro/","score_publie":0.709822,"score_0_1":0.7098,"niveau_suggere":82,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":9,"modele":"gpt-5-1-codex-max","nom":"GPT-5.1-Codex-Max","editeur":"OpenAI","page":"https://quelleia.com/modeles/gpt-5-1-codex-max/","score_publie":0.708301,"score_0_1":0.7083,"niveau_suggere":81.8,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/"},{"rang":10,"modele":"gpt-5","nom":"GPT-5","editeur":"OpenAI","page":"https://quelleia.com/modeles/gpt-5/","score_publie":0.696053,"score_0_1":0.6961,"niveau_suggere":80.5,"reflexion":"moyenne","harnais":null,"configurations_mesurees":2,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":11,"modele":"claude-sonnet-4-5","nom":"Claude Sonnet 4.5","editeur":"Anthropic","page":"https://quelleia.com/modeles/claude-sonnet-4-5/","score_publie":0.673842,"score_0_1":0.6738,"niveau_suggere":78.2,"reflexion":"budget","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":12,"modele":"claude-opus-4-1","nom":"Claude Opus 4.1","editeur":"Anthropic","page":"https://quelleia.com/modeles/claude-opus-4-1/","score_publie":0.668069,"score_0_1":0.6681,"niveau_suggere":77.6,"reflexion":"budget","harnais":null,"configurations_mesurees":2,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":13,"modele":"grok-4","nom":"Grok 4","editeur":"xAI","page":"https://quelleia.com/modeles/grok-4/","score_publie":0.665803,"score_0_1":0.6658,"niveau_suggere":77.4,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":14,"modele":"o3","nom":"o3","editeur":"OpenAI","page":"https://quelleia.com/modeles/o3/","score_publie":0.654376,"score_0_1":0.6544,"niveau_suggere":76.3,"reflexion":"moyenne","harnais":null,"configurations_mesurees":2,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":15,"modele":"claude-opus-4","nom":"Claude Opus 4","editeur":"Anthropic","page":"https://quelleia.com/modeles/claude-opus-4/","score_publie":0.639292,"score_0_1":0.6393,"niveau_suggere":74.8,"reflexion":"budget","harnais":null,"configurations_mesurees":2,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":16,"modele":"o4-mini","nom":"o4-mini","editeur":"OpenAI","page":"https://quelleia.com/modeles/o4-mini/","score_publie":0.639269,"score_0_1":0.6393,"niveau_suggere":74.8,"reflexion":"moyenne","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":17,"modele":"claude-sonnet-4","nom":"Claude Sonnet 4","editeur":"Anthropic","page":"https://quelleia.com/modeles/claude-sonnet-4/","score_publie":0.619725,"score_0_1":0.6197,"niveau_suggere":72.9,"reflexion":"budget","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":18,"modele":"claude-3-7-sonnet","nom":"Claude 3.7 Sonnet","editeur":"Anthropic","page":"https://quelleia.com/modeles/claude-3-7-sonnet/","score_publie":0.599681,"score_0_1":0.5997,"niveau_suggere":71,"reflexion":"budget","harnais":null,"configurations_mesurees":2,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":19,"modele":"kimi-k2-thinking","nom":"Kimi K2 Thinking","editeur":"Moonshot","page":"https://quelleia.com/modeles/kimi-k2-thinking/","score_publie":0.591943,"score_0_1":0.5919,"niveau_suggere":70.3,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":20,"modele":"gpt-oss-120b","nom":"gpt-oss-120b","editeur":"OpenAI","page":"https://quelleia.com/modeles/gpt-oss-120b/","score_publie":0.56634,"score_0_1":0.5663,"niveau_suggere":68,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":21,"modele":"gemini-2-5-pro-jun-2025","nom":"Gemini 2.5 Pro (Jun 2025)","editeur":"Google DeepMind","page":"https://quelleia.com/modeles/gemini-2-5-pro-jun-2025/","score_publie":0.5544,"score_0_1":0.5544,"niveau_suggere":66.9,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":22,"modele":"deepseek-r1-may-2025","nom":"DeepSeek-R1 (May 2025)","editeur":"DeepSeek","page":"https://quelleia.com/modeles/deepseek-r1-may-2025/","score_publie":0.537826,"score_0_1":0.5378,"niveau_suggere":65.4,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":23,"modele":"deepseek-r1","nom":"DeepSeek-R1","editeur":"DeepSeek","page":"https://quelleia.com/modeles/deepseek-r1/","score_publie":0.519251,"score_0_1":0.5193,"niveau_suggere":63.7,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":24,"modele":"o1","nom":"o1","editeur":"OpenAI","page":"https://quelleia.com/modeles/o1/","score_publie":0.510512,"score_0_1":0.5105,"niveau_suggere":63,"reflexion":"moyenne","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/"},{"rang":25,"modele":"deepseek-v3-mar-2025","nom":"DeepSeek-V3 (Mar 2025)","editeur":"DeepSeek","page":"https://quelleia.com/modeles/deepseek-v3-mar-2025/","score_publie":0.495813,"score_0_1":0.4958,"niveau_suggere":61.6,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":26,"modele":"deepseek-v3","nom":"DeepSeek-V3","editeur":"DeepSeek","page":"https://quelleia.com/modeles/deepseek-v3/","score_publie":0.473627,"score_0_1":0.4736,"niveau_suggere":59.7,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":27,"modele":"claude-3-5-sonnet-october-2024","nom":"Claude 3.5 Sonnet (October 2024)","editeur":"Anthropic","page":"https://quelleia.com/modeles/claude-3-5-sonnet-october-2024/","score_publie":0.452356,"score_0_1":0.4524,"niveau_suggere":57.7,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/"},{"rang":28,"modele":"o1-preview","nom":"o1-preview","editeur":"OpenAI","page":"https://quelleia.com/modeles/o1-preview/","score_publie":0.451441,"score_0_1":0.4514,"niveau_suggere":57.7,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/"},{"rang":29,"modele":"gpt-4o-nov-2024","nom":"GPT-4o (Nov 2024)","editeur":"OpenAI","page":"https://quelleia.com/modeles/gpt-4o-nov-2024/","score_publie":0.407636,"score_0_1":0.4076,"niveau_suggere":53.7,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":30,"modele":"claude-3-5-sonnet","nom":"Claude 3.5 Sonnet","editeur":"Anthropic","page":"https://quelleia.com/modeles/claude-3-5-sonnet/","score_publie":0.401502,"score_0_1":0.4015,"niveau_suggere":53.1,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/"},{"rang":31,"modele":"gpt-4-turbo-apr-2024","nom":"GPT-4 Turbo (Apr 2024)","editeur":"OpenAI","page":"https://quelleia.com/modeles/gpt-4-turbo-apr-2024/","score_publie":0.367357,"score_0_1":0.3674,"niveau_suggere":49.9,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":32,"modele":"gpt-4-mar-2023","nom":"GPT-4 (Mar 2023)","editeur":"OpenAI","page":"https://quelleia.com/modeles/gpt-4-mar-2023/","score_publie":0.361067,"score_0_1":0.3611,"niveau_suggere":49.3,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":33,"modele":"qwen2-5-72b","nom":"Qwen2.5-72B","editeur":"Alibaba","page":"https://quelleia.com/modeles/qwen2-5-72b/","score_publie":0.357771,"score_0_1":0.3578,"niveau_suggere":48.9,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":34,"modele":"gpt-4-turbo-nov-2023","nom":"GPT-4 Turbo (Nov 2023)","editeur":"OpenAI","page":"https://quelleia.com/modeles/gpt-4-turbo-nov-2023/","score_publie":0.351726,"score_0_1":0.3517,"niveau_suggere":48.4,"reflexion":"inconnue","harnais":null,"configurations_mesurees":3,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":35,"modele":"gpt-4o-aug-2024","nom":"GPT-4o (Aug 2024)","editeur":"OpenAI","page":"https://quelleia.com/modeles/gpt-4o-aug-2024/","score_publie":0.338424,"score_0_1":0.3384,"niveau_suggere":47,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":36,"modele":"qwen2-72b","nom":"Qwen2-72B","editeur":"Alibaba","page":"https://quelleia.com/modeles/qwen2-72b/","score_publie":0.298999,"score_0_1":0.299,"niveau_suggere":43,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":37,"modele":"claude-3-opus","nom":"Claude 3 Opus","editeur":"Anthropic","page":"https://quelleia.com/modeles/claude-3-opus/","score_publie":0.294753,"score_0_1":0.2948,"niveau_suggere":42.5,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/"},{"rang":38,"modele":"gpt-4-jun-2023","nom":"GPT-4 (Jun 2023)","editeur":"OpenAI","page":"https://quelleia.com/modeles/gpt-4-jun-2023/","score_publie":0.293044,"score_0_1":0.293,"niveau_suggere":42.3,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://epoch.ai/benchmarks"},{"rang":39,"modele":"gpt-3-5-turbo-instruct","nom":"GPT-3.5 Turbo Instruct","editeur":"OpenAI","page":"https://quelleia.com/modeles/gpt-3-5-turbo-instruct/","score_publie":0.214611,"score_0_1":0.2146,"niveau_suggere":33,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/"},{"rang":40,"modele":"gpt-3-175b-davinci","nom":"GPT-3 175B (davinci)","editeur":"OpenAI","page":"https://quelleia.com/modeles/gpt-3-175b-davinci/","score_publie":0.161869,"score_0_1":0.1619,"niveau_suggere":25.3,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/"},{"rang":41,"modele":"gpt-2-1-5b","nom":"GPT-2 (1.5B)","editeur":"OpenAI","page":"https://quelleia.com/modeles/gpt-2-1-5b/","score_publie":0.101046,"score_0_1":0.101,"niveau_suggere":13.2,"reflexion":"inconnue","harnais":null,"configurations_mesurees":1,"date":null,"source":"https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/"}]}