[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-post-\u002Fnews\u002Fanthropic-flashprove-provable-inference-which-model-answered":3,"news-related-en-\u002Fnews\u002Fanthropic-flashprove-provable-inference-which-model-answered":158,"hreflang-anthropic-flashprove-provable-inference-2026-11":175},{"doc":4,"redirect":8},{"id":5,"title":6,"author":7,"author_avatar":8,"body":9,"category":138,"date":139,"description":140,"extension":141,"image":142,"meta":143,"navigation":144,"path":145,"seo":146,"seoDescription":147,"seoTitle":148,"stem":149,"tags":150,"translationKey":156,"updatedAt":8,"__hash__":157},"news_en\u002Fnews\u002Fanthropic-flashprove-provable-inference-which-model-answered.md","Anthropic's FlashProve aims to prove which AI model actually answered you","Paul Lefizelier",null,{"type":10,"value":11,"toc":130},"minimark",[12,31,36,44,61,78,82,93,99,103,119,123],[13,14,15,16,24,25,30],"p",{},"When you pay for a given model through an API, you see only the tokens that come back. Nothing proves the provider did not quietly route you to a cheaper model. On 2 November 2026, Anthropic's Frontier Red Team ",[17,18,23],"a",{"href":19,"rel":20,"target":22},"https:\u002F\u002Fwww.anthropic.com\u002Fresearch\u002Fprovable-inference",[21],"nofollow","_blank","published progress on \"provable inference\""," and released ",[17,26,29],{"href":27,"rel":28,"target":22},"https:\u002F\u002Fgithub.com\u002Fanthropics\u002FFlashProve",[21],"FlashProve",", an open-source prototype meant to close that gap.",[32,33,35],"h2",{"id":34},"how-it-works","How it works",[13,37,38,39,43],{},"FlashProve has two phases (",[17,40,42],{"href":19,"rel":41,"target":22},[21],"Anthropic","):",[45,46,47,55],"ul",{},[48,49,50,54],"li",{},[51,52,53],"strong",{},"Offline:"," the provider publishes a commitment to the model weights, \"a short fingerprint\" that pins down one exact set of weights while keeping them hidden.",[48,56,57,60],{},[51,58,59],{},"Online:"," for each request, the provider produces a proof file showing that the committed model produces those output tokens.",[13,62,63,64,68,69,73,74,77],{},"Anyone holding the commitment, the prompt and the reply can check the proof without the weights. The ",[17,65,67],{"href":27,"rel":66,"target":22},[21],"GitHub repository"," says verification runs \"on an ordinary CPU in seconds\" with a tool called ",[70,71,72],"code",{},"cp_verify",". The system relies on zero-knowledge succinct proofs (zkSNARKs). It hides the weights, but hiding the model's architecture is not a goal (",[17,75,42],{"href":19,"rel":76,"target":22},[21],").",[32,79,81],{"id":80},"the-numbers","The numbers",[13,83,84,85,88,89,77],{},"Anthropic tested it on Llama 4 Maverick, a 400B-parameter mixture-of-experts model, using its public 4-bit weights. A single H100 GPU generates proofs at around 110 tokens per second, with a proof of about 19 MB per 1,024-token request. That is roughly 70 to 95 times the GPU time of running the model itself on vLLM (",[17,86,42],{"href":19,"rel":87,"target":22},[21],"). The repository lists a 1,024-token context window for now (",[17,90,92],{"href":27,"rel":91,"target":22},[21],"GitHub",[13,94,95,96,77],{},"Anthropic says Claude produced improvements to the proof system, and that the team countered hard-to-audit optimisation choices by stating the security guarantees in Lean, a formal verification language. Code, Lean proofs and documentation are released under Apache 2.0 (",[17,97,42],{"href":19,"rel":98,"target":22},[21],[32,100,102],{"id":101},"what-it-is-not","What it is not",[13,104,105,106,109,110,113,114,77],{},"This is not a product. Anthropic calls the post \"intermediate progress, and not a completed research agenda\" (",[17,107,42],{"href":19,"rel":108,"target":22},[21],"), the repository is marked as a research artifact that is not maintained (",[17,111,92],{"href":27,"rel":112,"target":22},[21],"), and TechCrunch reports there are no plans to deploy it at Anthropic or any other lab (",[17,115,118],{"href":116,"rel":117,"target":22},"https:\u002F\u002Ftechcrunch.com\u002F2026\u002F11\u002F03\u002Fanthropic-is-working-on-a-system-to-prove-ai-models-identity\u002F",[21],"TechCrunch",[32,120,122],{"id":121},"why-it-matters-for-ai-users","Why it matters for AI users",[13,124,125,126,129],{},"TechCrunch notes the researchers see it as a way to stop a provider from secretly swapping in a cheaper model, or a stronger one to inflate benchmark results (",[17,127,118],{"href":116,"rel":128,"target":22},[21],"). For anyone buying model access through resellers or routers, or paying for a \"frontier model reviewed this\" claim, a checkable proof would replace trust with verification. The 70 to 95x compute overhead and the short context window show it is still far from everyday use.",{"title":131,"searchDepth":132,"depth":132,"links":133},"",2,[134,135,136,137],{"id":34,"depth":132,"text":35},{"id":80,"depth":132,"text":81},{"id":101,"depth":132,"text":102},{"id":121,"depth":132,"text":122},"ai","2026-11-04T22:15:00","On 2 November 2026 Anthropic's Frontier Red Team released FlashProve, an open-source prototype that lets a provider prove a reply came from a committed model without revealing its weights. It is research code, with no deployment planned.","md","\u002Fimages\u002Fnews\u002Fanthropic-flashprove-provable-inference.jpg",{},true,"\u002Fnews\u002Fanthropic-flashprove-provable-inference-which-model-answered",{"title":6,"description":140},"Anthropic's FlashProve uses zero-knowledge proofs so anyone can check a reply came from a committed model, without the weights. Llama 4 Maverick, one H100, about 110 tokens\u002Fs.","FlashProve: Anthropic's open-source proof of which model answered","news\u002Fanthropic-flashprove-provable-inference-which-model-answered",[151,152,153,154,155],"anthropic","flashprove","cryptography","ai-safety","open-source","anthropic-flashprove-provable-inference-2026-11","DZ0HRWbeipwSLGA0spCdcHlIJHZrwUjIev3EWhjOeUY",[159,164,170],{"title":160,"description":161,"path":162,"date":139,"image":163,"category":138},"Crouton Alpha: a free, anonymous 1M-token model lands on OpenRouter","OpenRouter listed Crouton Alpha on 3 November 2026, a free stealth model for coding agents, tool use and document analysis with a 1,048,576-token context. Its developer is anonymous, and prompts may be retained.","\u002Fnews\u002Fopenrouter-crouton-alpha-free-stealth-model-1m-context","\u002Fimages\u002Fnews\u002Fopenrouter-crouton-alpha-stealth-model.jpg",{"title":165,"description":166,"path":167,"date":168,"image":169,"category":138},"EU court adviser: AI chatbots like Gemini can infringe press publishers' rights","On 3 November 2026, Advocate General Maciej Szpunar told the EU Court of Justice that a chatbot reusing news articles in its answers may infringe publishers' rights, while training on them is, as a rule, allowed.","\u002Fnews\u002Feu-court-adviser-gemini-chatbot-press-publishers-rights","2026-11-04T17:30:00","\u002Fimages\u002Fnews\u002Fcjeu-gemini-press-publishers-rights.jpg",{"title":171,"description":172,"path":173,"date":168,"image":174,"category":138},"OpenAI raises its bug bounty: up to $172,000 for the worst ChatGPT flaws","On 2 November 2026, OpenAI raised Security Bug Bounty rewards. Critical bugs in ChatGPT, Codex or the API Platform now pay up to $30,000, and exceptional findings up to $172,000 with new bonuses.","\u002Fnews\u002Fopenai-bug-bounty-172000-chatgpt-codex-api","\u002Fimages\u002Fnews\u002Fopenai-bug-bounty-172000.jpg",{"en":176,"fr":258,"es":8},{"id":5,"title":6,"author":7,"author_avatar":8,"body":177,"category":138,"date":139,"description":140,"extension":141,"image":142,"meta":255,"navigation":144,"path":145,"seo":256,"seoDescription":147,"seoTitle":148,"stem":149,"tags":257,"translationKey":156,"updatedAt":8,"__hash__":157},{"type":10,"value":178,"toc":249},[179,187,189,194,204,214,216,224,229,231,242,244],[13,180,15,181,24,184,30],{},[17,182,23],{"href":19,"rel":183,"target":22},[21],[17,185,29],{"href":27,"rel":186,"target":22},[21],[32,188,35],{"id":34},[13,190,38,191,43],{},[17,192,42],{"href":19,"rel":193,"target":22},[21],[45,195,196,200],{},[48,197,198,54],{},[51,199,53],{},[48,201,202,60],{},[51,203,59],{},[13,205,63,206,68,209,73,211,77],{},[17,207,67],{"href":27,"rel":208,"target":22},[21],[70,210,72],{},[17,212,42],{"href":19,"rel":213,"target":22},[21],[32,215,81],{"id":80},[13,217,84,218,88,221,77],{},[17,219,42],{"href":19,"rel":220,"target":22},[21],[17,222,92],{"href":27,"rel":223,"target":22},[21],[13,225,95,226,77],{},[17,227,42],{"href":19,"rel":228,"target":22},[21],[32,230,102],{"id":101},[13,232,105,233,109,236,113,239,77],{},[17,234,42],{"href":19,"rel":235,"target":22},[21],[17,237,92],{"href":27,"rel":238,"target":22},[21],[17,240,118],{"href":116,"rel":241,"target":22},[21],[32,243,122],{"id":121},[13,245,125,246,129],{},[17,247,118],{"href":116,"rel":248,"target":22},[21],{"title":131,"searchDepth":132,"depth":132,"links":250},[251,252,253,254],{"id":34,"depth":132,"text":35},{"id":80,"depth":132,"text":81},{"id":101,"depth":132,"text":102},{"id":121,"depth":132,"text":122},{},{"title":6,"description":140},[151,152,153,154,155],{"id":259,"title":260,"author":7,"author_avatar":8,"body":261,"category":138,"date":139,"description":369,"extension":141,"image":142,"meta":370,"navigation":144,"path":371,"seo":372,"seoDescription":373,"seoTitle":374,"stem":375,"tags":376,"translationKey":156,"updatedAt":8,"__hash__":379},"news_fr\u002Ffr\u002Fnews\u002Fanthropic-flashprove-prouver-quel-modele-ia-a-repondu.md","FlashProve : Anthropic veut prouver quel modèle d'IA vous a vraiment répondu",{"type":10,"value":262,"toc":363},[263,275,279,286,300,314,318,328,334,338,352,356],[13,264,265,266,270,271,274],{},"Quand vous payez un modèle précis via une API, vous ne voyez que les tokens renvoyés. Rien ne prouve que le fournisseur ne vous a pas discrètement basculé vers un modèle moins cher. Le 2 novembre 2026, la Frontier Red Team d'Anthropic a ",[17,267,269],{"href":19,"rel":268,"target":22},[21],"publié ses avancées sur l'« inférence prouvable »"," et sorti ",[17,272,29],{"href":27,"rel":273,"target":22},[21],", un prototype open source censé combler ce manque.",[32,276,278],{"id":277},"comment-ça-marche","Comment ça marche",[13,280,281,282,285],{},"FlashProve fonctionne en deux phases (",[17,283,42],{"href":19,"rel":284,"target":22},[21],") :",[45,287,288,294],{},[48,289,290,293],{},[51,291,292],{},"Hors ligne :"," le fournisseur publie un engagement sur les poids du modèle, une « courte empreinte » qui fixe un jeu de poids précis tout en le gardant secret.",[48,295,296,299],{},[51,297,298],{},"En ligne :"," pour chaque requête, le fournisseur produit un fichier de preuve montrant que le modèle engagé produit bien ces tokens de sortie.",[13,301,302,303,307,308,310,311,77],{},"Toute personne disposant de l'engagement, du prompt et de la réponse peut vérifier la preuve sans les poids. Le ",[17,304,306],{"href":27,"rel":305,"target":22},[21],"dépôt GitHub"," indique que la vérification tourne « sur un CPU ordinaire en quelques secondes » avec l'outil ",[70,309,72],{},". Le système repose sur des preuves succinctes à divulgation nulle de connaissance (zkSNARK). Il cache les poids, mais pas l'architecture du modèle (",[17,312,42],{"href":19,"rel":313,"target":22},[21],[32,315,317],{"id":316},"les-chiffres","Les chiffres",[13,319,320,321,324,325,77],{},"Anthropic l'a testé sur Llama 4 Maverick, un modèle mixture-of-experts de 400 milliards de paramètres, avec ses poids publics en 4 bits. Un seul GPU H100 génère des preuves à environ 110 tokens par seconde, pour une preuve d'environ 19 Mo par requête de 1 024 tokens. Cela représente environ 70 à 95 fois le temps GPU nécessaire pour faire tourner le modèle lui-même sur vLLM (",[17,322,42],{"href":19,"rel":323,"target":22},[21],"). Le dépôt indique une fenêtre de contexte de 1 024 tokens pour l'instant (",[17,326,92],{"href":27,"rel":327,"target":22},[21],[13,329,330,331,77],{},"Anthropic explique que Claude a apporté des améliorations au système de preuve, et que l'équipe a compensé des choix d'optimisation difficiles à auditer en énonçant les garanties de sécurité en Lean, un langage de vérification formelle. Le code, les preuves Lean et la documentation sont publiés sous licence Apache 2.0 (",[17,332,42],{"href":19,"rel":333,"target":22},[21],[32,335,337],{"id":336},"ce-que-ce-nest-pas","Ce que ce n'est pas",[13,339,340,341,344,345,348,349,77],{},"Ce n'est pas un produit. Anthropic parle de « progrès intermédiaires, pas d'un programme de recherche achevé » (",[17,342,42],{"href":19,"rel":343,"target":22},[21],"), le dépôt est présenté comme un artefact de recherche non maintenu (",[17,346,92],{"href":27,"rel":347,"target":22},[21],"), et TechCrunch rapporte qu'aucun déploiement n'est prévu chez Anthropic ni dans un autre laboratoire (",[17,350,118],{"href":116,"rel":351,"target":22},[21],[32,353,355],{"id":354},"pourquoi-cest-important-pour-les-utilisateurs-dia","Pourquoi c'est important pour les utilisateurs d'IA",[13,357,358,359,362],{},"Selon TechCrunch, les chercheurs y voient un moyen d'empêcher un fournisseur de remplacer en secret un modèle par un moins cher, ou par un plus puissant pour gonfler ses résultats de benchmark (",[17,360,118],{"href":116,"rel":361,"target":22},[21],"). Pour ceux qui achètent l'accès aux modèles via des revendeurs ou des routeurs, ou qui paient pour une affirmation du type « vérifié par un modèle de pointe », une preuve vérifiable remplacerait la confiance par un contrôle. Le surcoût de calcul de 70 à 95 fois et la fenêtre de contexte courte montrent que l'usage quotidien est encore loin.",{"title":131,"searchDepth":132,"depth":132,"links":364},[365,366,367,368],{"id":277,"depth":132,"text":278},{"id":316,"depth":132,"text":317},{"id":336,"depth":132,"text":337},{"id":354,"depth":132,"text":355},"Le 2 novembre 2026, la Frontier Red Team d'Anthropic a publié FlashProve, un prototype open source qui permet à un fournisseur de prouver qu'une réponse vient d'un modèle donné sans révéler ses poids. C'est du code de recherche, sans déploiement prévu.",{},"\u002Ffr\u002Fnews\u002Fanthropic-flashprove-prouver-quel-modele-ia-a-repondu",{"title":260,"description":369},"FlashProve d'Anthropic utilise des preuves à divulgation nulle pour vérifier qu'une réponse vient d'un modèle engagé, sans ses poids. Llama 4 Maverick, un H100, environ 110 tokens\u002Fs.","FlashProve : la preuve open source d'Anthropic sur le modèle utilisé","fr\u002Fnews\u002Fanthropic-flashprove-prouver-quel-modele-ia-a-repondu",[151,152,377,378,155],"cryptographie","securite-ia","E6H-X-2XihlGR4KfSx6xYbfPFfhzn-kp88r-Qu75y6A"]