[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-post-\u002Fnews\u002Fpentagon-8-classified-ai-contracts-openai-google-spacex-anthropic-excluded-may-2026":3,"news-related-en-\u002Fnews\u002Fpentagon-8-classified-ai-contracts-openai-google-spacex-anthropic-excluded-may-2026":633,"hreflang-pentagon-8-classified-ai-contracts-openai-google-spacex-anthropic-excluded-may-2026":653},{"doc":4,"redirect":8},{"id":5,"title":6,"author":7,"author_avatar":8,"body":9,"category":605,"date":606,"description":607,"extension":608,"image":609,"meta":610,"navigation":611,"path":612,"seo":613,"seoDescription":614,"seoTitle":615,"stem":616,"tags":617,"translationKey":631,"updatedAt":8,"__hash__":632},"news_en\u002Fnews\u002Fpentagon-8-classified-ai-contracts-openai-google-spacex-anthropic-excluded-may-2026.md","Pentagon: 8 classified AI contracts signed with OpenAI, Google, Microsoft, AWS, Nvidia, SpaceX, Oracle and Reflection — Anthropic permanently excluded","Paul Lefizelier",null,{"type":10,"value":11,"toc":583},"minimark",[12,64,69,80,108,115,119,238,253,257,267,306,313,317,320,325,336,340,357,361,368,372,389,400,404,416,432,436,439,443,456,460,467,471,483,487,490,519,531,535,550,561,564],[13,14,15,16,20,21,24,25,28,29,32,33,36,37,40,41,49,50,55,56,59,60,63],"p",{},"On ",[17,18,19],"strong",{},"May 1, 2026",", the ",[17,22,23],{},"Department of Defense"," (Pentagon) finalized a sweeping set of AI agreements for ",[17,26,27],{},"classified IL6 and IL7 networks"," with ",[17,30,31],{},"eight tech giants",": ",[17,34,35],{},"OpenAI, Google, Microsoft, Amazon Web Services, Nvidia, SpaceX, Oracle"," and the startup ",[17,38,39],{},"Reflection AI",". According to ",[42,43,48],"a",{"href":44,"rel":45,"target":47},"https:\u002F\u002Fwww.cnn.com\u002F2026\u002F05\u002F01\u002Ftech\u002Fpentagon-ai-anthropic",[46],"nofollow","_blank","CNN Business"," and ",[42,51,54],{"href":52,"rel":53,"target":47},"https:\u002F\u002Fbreakingdefense.com\u002F2026\u002F05\u002Fpentagon-clears-7-tech-firms-to-deploy-their-ai-on-its-classified-networks\u002F",[46],"Breaking Defense",", these new contracts replace the ",[17,57,58],{},"$200 million"," agreement signed with Anthropic in July 2025 — a contract that had made Anthropic the first AI lab to operate on Pentagon classified systems. ",[17,61,62],{},"Anthropic is permanently excluded",". The U.S. military AI market just flipped.",[65,66,68],"h2",{"id":67},"context-why-anthropic-was-ejected","Context: why Anthropic was ejected",[13,70,71,72,75,76,79],{},"To grasp the scale of the shift, recall the timeline. Anthropic was originally the Pentagon's ",[17,73,74],{},"reference AI provider",". The $200M contract from July 2025 confirmed that. But the breakdown accelerated in late 2025 over a very specific point: the ",[17,77,78],{},"usage clause",".",[13,81,82,83,86,87,49,90,93,94,98,99,102,103,107],{},"The Trump administration, through the DoD, wanted Claude to be ",[17,84,85],{},"usable \"for all lawful purposes\""," — including ",[17,88,89],{},"autonomous weapons",[17,91,92],{},"mass surveillance",". Anthropic refused. As we analyzed in ",[42,95,97],{"href":96},"\u002Fnews\u002Ftrump-anthropic-pentagon-ban-executive-order-2026","the March 2026 Anthropic Pentagon ban",", Dario Amodei's stance was consistent: Anthropic's ",[17,100,101],{},"Usage Policy"," explicitly prohibits using Claude for autonomous weapons systems or untargeted surveillance. There was a ",[42,104,106],{"href":105},"\u002Fnews\u002Ftrump-anthropic-reconciliation-pentagon-dario-amodei-mythos-april-2026","reconciliation attempt in April",", notably via the Mythos partnership, but it failed.",[13,109,110,111,114],{},"On May 1, 2026, the Pentagon ruled: ",[17,112,113],{},"the seven other tech giants (and Reflection AI) accept the clauses, so they win the contracts",". Anthropic doesn't sign, so loses everything.",[65,116,118],{"id":117},"the-eight-winners-who-signs-what","The eight winners: who signs what",[120,121,122,138],"table",{},[123,124,125],"thead",{},[126,127,128,132,135],"tr",{},[129,130,131],"th",{},"Company",[129,133,134],{},"Contract type",[129,136,137],{},"Classification level",[139,140,141,155,167,179,191,203,215,227],"tbody",{},[126,142,143,149,152],{},[144,145,146],"td",{},[17,147,148],{},"OpenAI",[144,150,151],{},"Frontier models, GPT-5.4 and Codex",[144,153,154],{},"IL6\u002FIL7",[126,156,157,162,165],{},[144,158,159],{},[17,160,161],{},"Google",[144,163,164],{},"Gemini Enterprise, A2A agents",[144,166,154],{},[126,168,169,174,177],{},[144,170,171],{},[17,172,173],{},"Microsoft",[144,175,176],{},"Foundry, Azure OpenAI Services",[144,178,154],{},[126,180,181,186,189],{},[144,182,183],{},[17,184,185],{},"Amazon AWS",[144,187,188],{},"Bedrock, Trainium 3 hosting",[144,190,154],{},[126,192,193,198,201],{},[144,194,195],{},[17,196,197],{},"Nvidia",[144,199,200],{},"NeMo, H300 GPU infra, NeMoClaw",[144,202,154],{},[126,204,205,210,213],{},[144,206,207],{},[17,208,209],{},"SpaceX",[144,211,212],{},"Starlink IL6\u002FIL7, Grok via xAI partnership",[144,214,154],{},[126,216,217,222,225],{},[144,218,219],{},[17,220,221],{},"Oracle",[144,223,224],{},"Cerner Defense, OCI Government",[144,226,154],{},[126,228,229,233,236],{},[144,230,231],{},[17,232,39],{},[144,234,235],{},"Proprietary frontier model, tactical agents",[144,237,154],{},[13,239,240,241,243,244,247,248,252],{},"Including ",[17,242,39],{}," is the surprise. This startup — founded by ex-DeepMind, valued at roughly $5B in early 2026 — enters the giants' courtyard directly. It's a sign the DoD wants to ",[17,245,246],{},"diversify away from incumbent big tech"," to keep negotiating leverage. As we noted in our ",[42,249,251],{"href":250},"\u002Fnews\u002Fnemoclaw-nvidia-enterprise-ai-agents-security","coverage of NeMoClaw at Nvidia",", Nvidia had paved the way for agents in classified environments since late 2025.",[65,254,256],{"id":255},"il6-and-il7-levels-what-it-really-means","IL6 and IL7 levels: what it really means",[13,258,259,262,263,266],{},[17,260,261],{},"Impact Level"," (IL) classifications, defined by the ",[17,264,265],{},"Defense Information Systems Agency (DISA)",", structure military AI deployment:",[268,269,270,277,283,289,298],"ul",{},[271,272,273,276],"li",{},[17,274,275],{},"IL2",": non-sensitive public data",[271,278,279,282],{},[17,280,281],{},"IL4",": controlled unclassified information (CUI)",[271,284,285,288],{},[17,286,287],{},"IL5",": higher-sensitivity CUI, mission-critical systems",[271,290,291,32,294,297],{},[17,292,293],{},"IL6",[17,295,296],{},"SECRET"," classified information",[271,299,300,32,303,297],{},[17,301,302],{},"IL7",[17,304,305],{},"TOP SECRET",[13,307,308,309,312],{},"The May 1 contracts grant IL6\u002FIL7 access. Concretely, that means ",[17,310,311],{},"Claude (Anthropic) won't run on Secret\u002FTS networks",", but GPT, Gemini, Grok, and Reflection's frontier model will. For AI researchers at OpenAI or Google, that's a massive business opportunity — but also a non-trivial ethical challenge. Several engineers have reportedly already pushed back internally, notably at Google (where the 2018 Project Maven episode led to resignations).",[65,314,316],{"id":315},"why-7-firms-accepted-what-anthropic-refused","Why 7 firms accepted what Anthropic refused",[13,318,319],{},"The ethical question is central. Why do seven giants sign a contract potentially covering autonomous weapons, where Anthropic refuses?",[321,322,324],"h3",{"id":323},"_1-a-different-reading-of-the-clauses","1. A different reading of the clauses",[13,326,327,328,331,332,335],{},"OpenAI, Google and Microsoft have ",[17,329,330],{},"more permissive Usage Policies"," than Anthropic, since an OpenAI revision in January 2024 removed the explicit ban on \"military and warfare applications.\" OpenAI's current wording leaves more interpretive room, with the commanding officer or operator responsible for avoiding prohibited uses case by case. Anthropic's stance is stricter: ",[17,333,334],{},"the policy serves as a technical envelope",", not just a moral guideline.",[321,337,339],{"id":338},"_2-revenue-pressure","2. Revenue pressure",[13,341,342,343,346,347,351,352,356],{},"At $200M per contract (likely more for the May 1 deals — some observers estimate ",[17,344,345],{},"$300-500M per firm","), the stakes are significant. For OpenAI, which has to fund its ",[42,348,350],{"href":349},"\u002Fnews\u002Fopenai-gpt-55-super-app-codex-atlas-agent-april-2026","GPT-5.5 super-app and the Codex Atlas agent",", this is revenue uncorrelated with consumer subscriptions. For Google, which is investing ",[42,353,355],{"href":354},"\u002Fnews\u002Fgoogle-anthropic-40-billion-investment-two-phase-cloud-tpu-april-2026","$40B in Anthropic",", it's strategic compensation: Google plays both sides.",[321,358,360],{"id":359},"_3-a-regulatory-precedent-effect","3. A regulatory precedent effect",[13,362,363,364,367],{},"With the new contracts, the DoD creates a ",[17,365,366],{},"U.S. legal precedent"," on military AI use. Signatories shape the rules. Anthropic, by staying out, loses that voice — a point openly criticized by several employees (internal rumors confirmed by Bloomberg).",[65,369,371],{"id":370},"anthropics-case-and-dario-amodeis-stance","Anthropic's case and Dario Amodei's stance",[13,373,374,375,378,379,383,384,49,387,79],{},"Dario Amodei has publicly reaffirmed his position: ",[17,376,377],{},"Anthropic's mission is to build safe AI, and certain military uses are incompatible with that mission",". That's a stance the safety-research community also defends, and which we already covered in our ",[42,380,382],{"href":381},"\u002Fnews\u002Fanthropic-project-glasswing-claude-mythos-strategic-manipulation-cybersecurity","Project Glasswing and Mythos analysis",". Anthropic isn't anti-defense — the firm keeps civilian government customers — but refuses ",[17,385,386],{},"autonomous weapons systems",[17,388,92],{},[13,390,391,392,395,396,399],{},"That position has a measurable business cost. According to internal numbers, the lost DoD pipeline is on the order of ",[17,393,394],{},"$3-5 billion over 5 years",". To put it in perspective with the ",[17,397,398],{},"$40B ARR run rate"," Anthropic estimates, that's marginal — but not zero.",[65,401,403],{"id":402},"reflection-ai-the-outsider-that-grabs-an-il6il7","Reflection AI: the outsider that grabs an IL6\u002FIL7",[13,405,406,407,410,411,415],{},"Reflection AI's entry into the classified-provider club is probably the most surprising development. Reflection was founded in late 2024 by ex-DeepMind staff with the ambition of building a frontier model focused on ",[17,408,409],{},"agentic reasoning for high-responsibility environments"," — finance, healthcare, defense. Their Series C (estimated at $800M at $5B valuation, early 2026) echoes the ",[42,412,414],{"href":413},"\u002Fnews\u002Fcursor-2-billion-50-billion-valuation-a16z-nvidia-vibe-coding-april-2026","Cursor round at $50B",": U.S. VCs are aggressively pushing on AI defense.",[13,417,418,419,49,423,427,428,431],{},"It's also a geopolitical signal. With growing concerns about AI sovereignty against China (cf. ",[42,420,422],{"href":421},"\u002Fnews\u002Fdeepseek-v4-open-weight-ai-model-1-trillion-parameters","DeepSeek V4 raising at $1T",[42,424,426],{"href":425},"\u002Fnews\u002Fdeepseek-tencent-alibaba-20-billion-valuation-first-round-april-2026","Tencent\u002FAlibaba at $20B","), the U.S. secures its military AI supply chain by including a ",[17,429,430],{},"domestic outsider"," (Reflection) alongside big tech.",[65,433,435],{"id":434},"geopolitical-implications","Geopolitical implications",[13,437,438],{},"The shift goes beyond the U.S. Three consequences to track:",[321,440,442],{"id":441},"_1-europe-positions-itself","1. Europe positions itself",[13,444,445,446,450,451,455],{},"With ",[42,447,449],{"href":448},"\u002Fnews\u002Fuk-sovereign-ai-fund-500-million-callosum-british-sovereignty-april-2026","the UK Sovereign AI Fund of £500M"," and the French push for ",[42,452,454],{"href":453},"\u002Fnews\u002Ffrench-parliament-big-tech-digital-sovereignty-inquiry-2026","a digital sovereignty inquiry commission",", Europe accelerates its own sovereignty strategy. Mistral, AMI Labs (Yann LeCun), and likely BAE Systems \u002F Thales are opening AI defense programs.",[321,457,459],{"id":458},"_2-china-and-russia-rearm","2. China and Russia rearm",[13,461,462,463,466],{},"Excluding Anthropic and signing massively with seven U.S. giants sends a clear signal to Beijing and Moscow: ",[17,464,465],{},"the U.S. integrates AI into its military chain",". Expected response: acceleration of the PLA's Tiangong program and faster deployment of Yandex \u002F Sber Cloud on Russian systems.",[321,468,470],{"id":469},"_3-defense-ai-startups-boom","3. Defense AI startups boom",[13,472,473,474,477,478,482],{},"Including Reflection paves the way for a ",[17,475,476],{},"second tier"," of defense AI startups. ",[42,479,481],{"href":480},"\u002Fnews\u002Fshield-ai-2-billion-series-g-aechelon-hivemind-defense-ai-april-2026","Shield AI already raised $2B Series G",", Anduril keeps expanding, and new entrants (Scout AI, Helsing, Hadrian) raise massive Series A rounds.",[65,484,486],{"id":485},"what-it-changes-for-ai-developers-and-consumer-apps","What it changes for AI developers and consumer apps",[13,488,489],{},"For developers and consumer apps, the direct impact is limited — but the global context shifts:",[268,491,492,498,504],{},[271,493,494,497],{},[17,495,496],{},"Public models stay the same",". GPT-5.4, Claude Opus 4.7, Gemini 2.5 Pro remain available via standard API",[271,499,500,503],{},[17,501,502],{},"Tech competition intensifies",". With DoD contracts at $300-500M, OpenAI and Google get extra funding leverage to push the frontier (GPT-6, Gemini 3.5)",[271,505,506,509,510,514,515,518],{},[17,507,508],{},"Talent arbitrage shifts",". As we analyzed in the ",[42,511,513],{"href":512},"\u002Fnews\u002Fai-brain-drain-meta-google-openai-startups-188-billion-vc-april-2026","Meta-Google brain drain to startups",", some safety researchers who don't want to work on defense will ",[17,516,517],{},"migrate to Anthropic"," or new structures",[13,520,521,522,525,526,530],{},"For consumer AI app publishers, the message stays consistent: AI product monetization runs through ",[17,523,524],{},"civilian users, not the DoD",". The chat SDK and native ad formats like ",[42,527,529],{"href":528},"\u002Fpublishers","Idlen's"," remain the fastest path to turn an AI app into a profitable product, without depending on government procurement cycles.",[65,532,534],{"id":533},"conclusion-the-ethical-fracture-in-us-ai-is-now-consummated","Conclusion: the ethical fracture in U.S. AI is now consummated",[13,536,537,538,541,542,545,546,549],{},"May 1, 2026 marks an ",[17,539,540],{},"ethical inflection point"," in the U.S. AI industry. For the first time since Project Maven in 2018, ",[17,543,544],{},"the seven largest U.S. AI labs"," (Anthropic excluded) have collectively decided that ",[17,547,548],{},"classified military contracts are worth bending the Usage Policies",". That's a precedent that will shape the next 10 years.",[13,551,552,553,556,557,79],{},"Anthropic, by staying out, pays a business price. But gains an ethical positioning that could, over time, become a massive commercial advantage in healthcare, education, and European customers sensitive to sovereignty. As Dario Amodei wrote in his latest internal memo: ",[17,554,555],{},"\"Our brand is our policy.\""," The next months will tell whether that discipline pays off in IPO valuation — a decisive topic for ",[42,558,560],{"href":559},"\u002Fnews\u002Fanthropic-900-billion-valuation-ipo-october-2026-50-billion-round-may-2026","Anthropic's path to an October 2026 IPO",[562,563],"hr",{},[13,565,566],{},[567,568,569,570,574,575,578,579,79],"em",{},"To continue on military and geopolitical AI stakes, see our coverage of ",[42,571,573],{"href":572},"\u002Fnews\u002Fpentagon-100000-ai-agents-vibe-coding-genai-mil-google-gemini-april-2026","100,000 AI agents deployed at the Pentagon",", our breakdown of the ",[42,576,577],{"href":96},"initial Trump-Anthropic ban"," and our ",[42,580,582],{"href":581},"\u002Fblog\u002Fhow-to-monetize-ai-app-complete-guide","complete guide to monetizing AI apps",{"title":584,"searchDepth":585,"depth":585,"links":586},"",2,[587,588,589,590,596,597,598,603,604],{"id":67,"depth":585,"text":68},{"id":117,"depth":585,"text":118},{"id":255,"depth":585,"text":256},{"id":315,"depth":585,"text":316,"children":591},[592,594,595],{"id":323,"depth":593,"text":324},3,{"id":338,"depth":593,"text":339},{"id":359,"depth":593,"text":360},{"id":370,"depth":585,"text":371},{"id":402,"depth":585,"text":403},{"id":434,"depth":585,"text":435,"children":599},[600,601,602],{"id":441,"depth":593,"text":442},{"id":458,"depth":593,"text":459},{"id":469,"depth":593,"text":470},{"id":485,"depth":585,"text":486},{"id":533,"depth":585,"text":534},"ai","2026-05-05T10:30:00","On May 1, 2026, the Pentagon finalized AI agreements for IL6\u002FIL7 classified networks with eight tech giants, deliberately excluding Anthropic. The $200M Anthropic contract is dead. Full breakdown of the military, geopolitical and ethical stakes.","md","\u002Fpentagon-8-classified-ai-contracts-openai-google-spacex-anthropic-excluded-may-2026.jpg",{},true,"\u002Fnews\u002Fpentagon-8-classified-ai-contracts-openai-google-spacex-anthropic-excluded-may-2026",{"title":6,"description":607},"Pentagon signs with OpenAI, Google, Microsoft, AWS, Nvidia, SpaceX, Oracle and Reflection. Anthropic excluded after refusing to amend its safety limits.","Pentagon 8 classified AI contracts May 2026: Anthropic out, OpenAI in","news\u002Fpentagon-8-classified-ai-contracts-openai-google-spacex-anthropic-excluded-may-2026",[618,619,620,621,622,623,624,625,626,627,628,629,630],"pentagon","department-of-defense","openai","google","microsoft","aws","nvidia","spacex","oracle","reflection-ai","anthropic","dod","ai-policy","pentagon-8-classified-ai-contracts-openai-google-spacex-anthropic-excluded-may-2026","wTza5Yb7TsCcEzQ0yC4oobZyNF9xRmmYfAWscaptGWA",[634,641,647],{"title":635,"description":636,"path":637,"date":638,"image":639,"category":640},"Cursor cloud agents now build, host and publish live apps","Cursor's cloud agents can scaffold a web app, provision Postgres and Cursor SSO, run security checks, then publish a shareable URL — without a separate hosting account.","\u002Fnews\u002Fcursor-live-apps-cloud-agents-deploy-postgres-auth","2026-11-07T08:00:00","\u002Fimages\u002Fnews\u002Fcursor-live-apps-cloud-agents-deploy-postgres-auth.jpg","vibecoding",{"title":642,"description":643,"path":644,"date":645,"image":646,"category":605},"Google makes Gemini Omni 1.2 Flash generally available with 15-second video","Google DeepMind cut Omni Flash pricing by 30% to $0.07 per second at 720p, raised clip length from 10 to 15 seconds, and opened the model on the Gemini API and Google Cloud.","\u002Fnews\u002Fgoogle-gemini-omni-1-2-flash-15-second-video","2026-11-06T08:00:00","\u002Fimages\u002Fnews\u002Fgoogle-gemini-omni-1-2-flash-15-second-video.jpg",{"title":648,"description":649,"path":650,"date":651,"image":652,"category":605},"Anthropic reportedly signs its first publisher licensing deal, with the AP","Adweek reports that Anthropic struck a content and data licensing deal with the Associated Press just before the US midterms, its first with a digital publisher. Neither company has confirmed it on the record, and the terms are undisclosed.","\u002Fnews\u002Fanthropic-first-publisher-licensing-deal-associated-press","2026-11-05T12:35:00","\u002Fimages\u002Fnews\u002Fanthropic-ap-publisher-licensing-deal.jpg",{"en":654,"fr":1015,"es":8},{"id":5,"title":6,"author":7,"author_avatar":8,"body":655,"category":605,"date":606,"description":607,"extension":608,"image":609,"meta":1012,"navigation":611,"path":612,"seo":1013,"seoDescription":614,"seoTitle":615,"stem":616,"tags":1014,"translationKey":631,"updatedAt":8,"__hash__":632},{"type":10,"value":656,"toc":993},[657,681,683,689,703,707,709,803,811,813,819,845,849,851,853,855,861,863,871,873,877,879,889,895,897,903,911,913,915,917,923,925,929,931,937,939,941,959,965,967,975,981,983],[13,658,15,659,20,661,24,663,28,665,32,667,36,669,40,671,49,674,55,677,59,679,63],{},[17,660,19],{},[17,662,23],{},[17,664,27],{},[17,666,31],{},[17,668,35],{},[17,670,39],{},[42,672,48],{"href":44,"rel":673,"target":47},[46],[42,675,54],{"href":52,"rel":676,"target":47},[46],[17,678,58],{},[17,680,62],{},[65,682,68],{"id":67},[13,684,71,685,75,687,79],{},[17,686,74],{},[17,688,78],{},[13,690,82,691,86,693,49,695,93,697,98,699,102,701,107],{},[17,692,85],{},[17,694,89],{},[17,696,92],{},[42,698,97],{"href":96},[17,700,101],{},[42,702,106],{"href":105},[13,704,110,705,114],{},[17,706,113],{},[65,708,118],{"id":117},[120,710,711,721],{},[123,712,713],{},[126,714,715,717,719],{},[129,716,131],{},[129,718,134],{},[129,720,137],{},[139,722,723,733,743,753,763,773,783,793],{},[126,724,725,729,731],{},[144,726,727],{},[17,728,148],{},[144,730,151],{},[144,732,154],{},[126,734,735,739,741],{},[144,736,737],{},[17,738,161],{},[144,740,164],{},[144,742,154],{},[126,744,745,749,751],{},[144,746,747],{},[17,748,173],{},[144,750,176],{},[144,752,154],{},[126,754,755,759,761],{},[144,756,757],{},[17,758,185],{},[144,760,188],{},[144,762,154],{},[126,764,765,769,771],{},[144,766,767],{},[17,768,197],{},[144,770,200],{},[144,772,154],{},[126,774,775,779,781],{},[144,776,777],{},[17,778,209],{},[144,780,212],{},[144,782,154],{},[126,784,785,789,791],{},[144,786,787],{},[17,788,221],{},[144,790,224],{},[144,792,154],{},[126,794,795,799,801],{},[144,796,797],{},[17,798,39],{},[144,800,235],{},[144,802,154],{},[13,804,240,805,243,807,247,809,252],{},[17,806,39],{},[17,808,246],{},[42,810,251],{"href":250},[65,812,256],{"id":255},[13,814,815,262,817,266],{},[17,816,261],{},[17,818,265],{},[268,820,821,825,829,833,839],{},[271,822,823,276],{},[17,824,275],{},[271,826,827,282],{},[17,828,281],{},[271,830,831,288],{},[17,832,287],{},[271,834,835,32,837,297],{},[17,836,293],{},[17,838,296],{},[271,840,841,32,843,297],{},[17,842,302],{},[17,844,305],{},[13,846,308,847,312],{},[17,848,311],{},[65,850,316],{"id":315},[13,852,319],{},[321,854,324],{"id":323},[13,856,327,857,331,859,335],{},[17,858,330],{},[17,860,334],{},[321,862,339],{"id":338},[13,864,342,865,346,867,351,869,356],{},[17,866,345],{},[42,868,350],{"href":349},[42,870,355],{"href":354},[321,872,360],{"id":359},[13,874,363,875,367],{},[17,876,366],{},[65,878,371],{"id":370},[13,880,374,881,378,883,383,885,49,887,79],{},[17,882,377],{},[42,884,382],{"href":381},[17,886,386],{},[17,888,92],{},[13,890,391,891,395,893,399],{},[17,892,394],{},[17,894,398],{},[65,896,403],{"id":402},[13,898,406,899,410,901,415],{},[17,900,409],{},[42,902,414],{"href":413},[13,904,418,905,49,907,427,909,431],{},[42,906,422],{"href":421},[42,908,426],{"href":425},[17,910,430],{},[65,912,435],{"id":434},[13,914,438],{},[321,916,442],{"id":441},[13,918,445,919,450,921,455],{},[42,920,449],{"href":448},[42,922,454],{"href":453},[321,924,459],{"id":458},[13,926,462,927,466],{},[17,928,465],{},[321,930,470],{"id":469},[13,932,473,933,477,935,482],{},[17,934,476],{},[42,936,481],{"href":480},[65,938,486],{"id":485},[13,940,489],{},[268,942,943,947,951],{},[271,944,945,497],{},[17,946,496],{},[271,948,949,503],{},[17,950,502],{},[271,952,953,509,955,514,957,518],{},[17,954,508],{},[42,956,513],{"href":512},[17,958,517],{},[13,960,521,961,525,963,530],{},[17,962,524],{},[42,964,529],{"href":528},[65,966,534],{"id":533},[13,968,537,969,541,971,545,973,549],{},[17,970,540],{},[17,972,544],{},[17,974,548],{},[13,976,552,977,556,979,79],{},[17,978,555],{},[42,980,560],{"href":559},[562,982],{},[13,984,985],{},[567,986,569,987,574,989,578,991,79],{},[42,988,573],{"href":572},[42,990,577],{"href":96},[42,992,582],{"href":581},{"title":584,"searchDepth":585,"depth":585,"links":994},[995,996,997,998,1003,1004,1005,1010,1011],{"id":67,"depth":585,"text":68},{"id":117,"depth":585,"text":118},{"id":255,"depth":585,"text":256},{"id":315,"depth":585,"text":316,"children":999},[1000,1001,1002],{"id":323,"depth":593,"text":324},{"id":338,"depth":593,"text":339},{"id":359,"depth":593,"text":360},{"id":370,"depth":585,"text":371},{"id":402,"depth":585,"text":403},{"id":434,"depth":585,"text":435,"children":1006},[1007,1008,1009],{"id":441,"depth":593,"text":442},{"id":458,"depth":593,"text":459},{"id":469,"depth":593,"text":470},{"id":485,"depth":585,"text":486},{"id":533,"depth":585,"text":534},{},{"title":6,"description":607},[618,619,620,621,622,623,624,625,626,627,628,629,630],{"id":1016,"title":1017,"author":7,"author_avatar":8,"body":1018,"category":605,"date":606,"description":1568,"extension":608,"image":609,"meta":1569,"navigation":611,"path":1570,"seo":1571,"seoDescription":1572,"seoTitle":1573,"stem":1574,"tags":1575,"translationKey":631,"updatedAt":8,"__hash__":1577},"news_fr\u002Ffr\u002Fnews\u002Fpentagone-8-contrats-ia-classifies-openai-google-spacex-anthropic-exclu-mai-2026.md","Pentagone : 8 contrats IA classifiés signés avec OpenAI, Google, Microsoft, AWS, Nvidia, SpaceX, Oracle et Reflection — Anthropic définitivement exclu",{"type":10,"value":1019,"toc":1549},[1020,1060,1064,1074,1102,1109,1113,1214,1229,1233,1242,1272,1279,1283,1286,1290,1301,1305,1322,1326,1333,1337,1354,1365,1369,1384,1400,1404,1407,1411,1424,1428,1435,1439,1451,1455,1458,1487,1499,1503,1518,1529,1531],[13,1021,1022,1023,1026,1027,1029,1030,1033,1034,1037,1038,1040,1041,1043,1044,1047,1048,1051,1052,1055,1056,1059],{},"Le ",[17,1024,1025],{},"1er mai 2026",", le ",[17,1028,23],{}," (Pentagone) a finalisé un ensemble d'accords IA pour les ",[17,1031,1032],{},"réseaux classifiés IL6 et IL7"," avec ",[17,1035,1036],{},"huit géants technologiques"," : ",[17,1039,35],{}," et la startup ",[17,1042,39],{},". Selon ",[42,1045,48],{"href":44,"rel":1046,"target":47},[46]," et ",[42,1049,54],{"href":52,"rel":1050,"target":47},[46],", ces nouveaux contrats remplacent l'",[17,1053,1054],{},"accord de 200 millions de dollars"," signé avec Anthropic en juillet 2025 — un contrat qui faisait alors d'Anthropic la première AI lab à travailler sur des systèmes classifiés du Pentagone. ",[17,1057,1058],{},"Anthropic est définitivement exclu",". Le marché militaire IA américain vient de basculer.",[65,1061,1063],{"id":1062},"le-contexte-pourquoi-anthropic-a-été-éjecté","Le contexte : pourquoi Anthropic a été éjecté",[13,1065,1066,1067,1070,1071,79],{},"Pour comprendre l'ampleur du basculement, il faut rappeler le timeline. Anthropic était initialement le ",[17,1068,1069],{},"fournisseur de référence"," du Pentagone sur l'IA. Le contrat de 200 M$ de juillet 2025 le confirmait. Mais la rupture s'est accélérée à partir de fin 2025 sur un point très précis : la ",[17,1072,1073],{},"clause d'usage",[13,1075,1076,1077,1080,1081,1084,1085,1088,1089,1093,1094,1096,1097,1101],{},"L'administration Trump, via le DoD, voulait que Claude soit ",[17,1078,1079],{},"utilisable « pour toutes les fins légales »"," — y compris les ",[17,1082,1083],{},"armes autonomes"," et la ",[17,1086,1087],{},"surveillance de masse",". Anthropic a refusé. Comme on l'avait analysé dans ",[42,1090,1092],{"href":1091},"\u002Ffr\u002Fnews\u002Ftrump-anthropic-ban-pentagone-decret-2026","le ban Pentagone d'Anthropic en mars 2026",", la position de Dario Amodei a été cohérente : la ",[17,1095,101],{}," d'Anthropic interdit explicitement l'usage de Claude pour des armes autonomes ou des systèmes de surveillance non-ciblés. Il y a eu une ",[42,1098,1100],{"href":1099},"\u002Ffr\u002Fnews\u002Ftrump-anthropic-reconciliation-pentagone-dario-amodei-mythos-avril-2026","tentative de réconciliation en avril",", notamment via le partenariat Mythos, mais elle a échoué.",[13,1103,1104,1105,1108],{},"Le 1er mai 2026, le Pentagone tranche : ",[17,1106,1107],{},"les sept autres géants tech (et Reflection AI) acceptent les clauses, donc gagnent les contrats",". Anthropic ne signe pas, donc perd tout.",[65,1110,1112],{"id":1111},"les-huit-gagnants-qui-signe-quoi","Les huit gagnants : qui signe quoi",[120,1114,1115,1128],{},[123,1116,1117],{},[126,1118,1119,1122,1125],{},[129,1120,1121],{},"Entreprise",[129,1123,1124],{},"Type de contrat",[129,1126,1127],{},"Niveau classification",[139,1129,1130,1141,1152,1162,1172,1183,1193,1203],{},[126,1131,1132,1136,1139],{},[144,1133,1134],{},[17,1135,148],{},[144,1137,1138],{},"Modèles frontier, GPT-5.4 et Codex",[144,1140,154],{},[126,1142,1143,1147,1150],{},[144,1144,1145],{},[17,1146,161],{},[144,1148,1149],{},"Gemini Enterprise, agents A2A",[144,1151,154],{},[126,1153,1154,1158,1160],{},[144,1155,1156],{},[17,1157,173],{},[144,1159,176],{},[144,1161,154],{},[126,1163,1164,1168,1170],{},[144,1165,1166],{},[17,1167,185],{},[144,1169,188],{},[144,1171,154],{},[126,1173,1174,1178,1181],{},[144,1175,1176],{},[17,1177,197],{},[144,1179,1180],{},"NeMo, infra GPU H300, NeMoClaw",[144,1182,154],{},[126,1184,1185,1189,1191],{},[144,1186,1187],{},[17,1188,209],{},[144,1190,212],{},[144,1192,154],{},[126,1194,1195,1199,1201],{},[144,1196,1197],{},[17,1198,221],{},[144,1200,224],{},[144,1202,154],{},[126,1204,1205,1209,1212],{},[144,1206,1207],{},[17,1208,39],{},[144,1210,1211],{},"Modèle frontier propriétaire, agents tactiques",[144,1213,154],{},[13,1215,1216,1217,1219,1220,1223,1224,1228],{},"L'inclusion de ",[17,1218,39],{}," est la surprise. Cette startup — fondée par d'anciens DeepMind, valorisée à environ 5 milliards début 2026 — entre directement dans la cour des géants. C'est le signe que le DoD veut ",[17,1221,1222],{},"diversifier hors des big tech"," historiques pour avoir des leviers de négociation. Comme le rappelait notre ",[42,1225,1227],{"href":1226},"\u002Ffr\u002Fnews\u002Fnemoclaw-nvidia-agents-ia-securite-enterprise","coverage de NeMoClaw chez Nvidia",", Nvidia avait préparé l'arrivée des agents en environnement classifié dès fin 2025.",[65,1230,1232],{"id":1231},"les-niveaux-il6-et-il7-ce-que-ça-signifie-vraiment","Les niveaux IL6 et IL7 : ce que ça signifie vraiment",[13,1234,1235,1236,1238,1239,1241],{},"Les niveaux d'",[17,1237,261],{}," définis par la ",[17,1240,265],{}," structurent le déploiement IA militaire :",[268,1243,1244,1249,1254,1259,1266],{},[271,1245,1246,1248],{},[17,1247,275],{}," : données publiques non-sensibles",[271,1250,1251,1253],{},[17,1252,281],{}," : controlled unclassified information (CUI)",[271,1255,1256,1258],{},[17,1257,287],{}," : higher sensitivity CUI, mission-critical systems",[271,1260,1261,1263,1264],{},[17,1262,293],{}," : informations classifiées ",[17,1265,296],{},[271,1267,1268,1263,1270],{},[17,1269,302],{},[17,1271,305],{},[13,1273,1274,1275,1278],{},"Les contrats du 1er mai donnent accès à IL6\u002FIL7. Concrètement, cela signifie que ",[17,1276,1277],{},"Claude (Anthropic) ne tournera pas sur les réseaux Secret\u002FTS",", mais que GPT, Gemini, Grok, ainsi que Reflection's frontier model, le pourront. Pour les chercheurs IA chez OpenAI ou Google, c'est une opportunité business massive — mais aussi un défi éthique non-trivial. Plusieurs ingénieurs auraient déjà manifesté en interne, notamment chez Google (où le précédent Project Maven en 2018 avait conduit à des démissions).",[65,1280,1282],{"id":1281},"pourquoi-7-entreprises-ont-accepté-ce-quanthropic-a-refusé","Pourquoi 7 entreprises ont accepté ce qu'Anthropic a refusé",[13,1284,1285],{},"La question éthique est centrale. Pourquoi sept géants signent un contrat qui inclut potentiellement des armes autonomes, là où Anthropic refuse ?",[321,1287,1289],{"id":1288},"_1-une-lecture-différente-des-clauses","1. Une lecture différente des clauses",[13,1291,1292,1293,1296,1297,1300],{},"OpenAI, Google et Microsoft ont des ",[17,1294,1295],{},"Usage Policies plus permissives"," que celle d'Anthropic, depuis une révision OpenAI en janvier 2024 qui avait retiré l'interdiction explicite des « military and warfare applications ». La formulation actuelle d'OpenAI laisse plus de marge d'interprétation, à charge du commanding officer ou de l'opérateur d'éviter les usages prohibés au cas par cas. La position d'Anthropic est plus stricte : ",[17,1298,1299],{},"la policy sert d'enveloppe technique",", pas seulement de guideline morale.",[321,1302,1304],{"id":1303},"_2-la-pression-revenue","2. La pression revenue",[13,1306,1307,1308,1311,1312,1316,1317,1321],{},"À 200 M$ par contrat (probablement plus pour les contrats du 1er mai, certains observateurs évoquant ",[17,1309,1310],{},"300-500 M$ par firme","), les sommes en jeu sont significatives. Pour OpenAI qui doit financer son ",[42,1313,1315],{"href":1314},"\u002Ffr\u002Fnews\u002Fopenai-gpt-55-super-app-codex-atlas-agent-avril-2026","super-app GPT-5.5 et son agent Codex Atlas",", c'est un revenu non-corrélé aux subscriptions consumer. Pour Google qui investit ",[42,1318,1320],{"href":1319},"\u002Ffr\u002Fnews\u002Fgoogle-anthropic-40-milliards-investissement-deux-phases-cloud-tpu-avril-2026","40 milliards dans Anthropic",", c'est une compensation stratégique : Google joue les deux côtés.",[321,1323,1325],{"id":1324},"_3-un-effet-de-précédent-réglementaire","3. Un effet de précédent réglementaire",[13,1327,1328,1329,1332],{},"Avec les nouveaux contrats, le DoD crée un ",[17,1330,1331],{},"précédent juridique américain"," sur l'usage IA militaire. Les firmes signataires participent à la définition des règles. Anthropic, en restant à l'écart, perd cette voix — un point que critiquent ouvertement plusieurs employés (rumeurs internes confirmées par Bloomberg).",[65,1334,1336],{"id":1335},"les-arguments-danthropic-et-la-position-de-dario-amodei","Les arguments d'Anthropic et la position de Dario Amodei",[13,1338,1339,1340,1343,1344,1348,1349,1084,1352,79],{},"Dario Amodei a publiquement réaffirmé sa position : ",[17,1341,1342],{},"la mission d'Anthropic est de construire une IA sûre, et certains usages militaires sont incompatibles avec cette mission",". C'est une position que défend également la communauté safety-research, et qu'on retrouvait déjà dans ",[42,1345,1347],{"href":1346},"\u002Ffr\u002Fnews\u002Fanthropic-project-glasswing-claude-mythos-strategic-manipulation-cybersecurite","notre analyse de Project Glasswing et Mythos",". Anthropic n'est pas anti-défense — la firme garde des clients gouvernementaux civils — mais refuse les ",[17,1350,1351],{},"systèmes d'armes autonomes",[17,1353,1087],{},[13,1355,1356,1357,1360,1361,1364],{},"Cette position a un coût business mesurable. Selon les chiffres internes, le manque à gagner sur le pipeline DoD serait de l'ordre de ",[17,1358,1359],{},"3 à 5 milliards de dollars sur 5 ans",". À mettre en perspective avec les ",[17,1362,1363],{},"40 milliards d'ARR run rate"," estimés par Anthropic, c'est marginal — mais pas zéro.",[65,1366,1368],{"id":1367},"reflection-ai-loutsider-qui-rafle-un-il6il7","Reflection AI : l'outsider qui rafle un IL6\u002FIL7",[13,1370,1371,1372,1374,1375,1378,1379,1383],{},"L'entrée de ",[17,1373,39],{}," dans le club des fournisseurs classifiés est probablement le développement le plus surprenant. Reflection a été fondée fin 2024 par des anciens DeepMind, avec l'ambition de construire un modèle frontier orienté ",[17,1376,1377],{},"reasoning agentique pour environnements à haute responsabilité"," — finance, santé, défense. La levée de leur Série C (estimée à 800 M$ à 5 Md$ de valorisation, début 2026) fait écho au ",[42,1380,1382],{"href":1381},"\u002Ffr\u002Fnews\u002Fcursor-2-milliards-50-milliards-valorisation-a16z-nvidia-vibe-coding-avril-2026","tour Cursor à 50 milliards"," : les VC américains poussent agressivement sur la défense IA.",[13,1385,1386,1387,1047,1391,1395,1396,1399],{},"C'est aussi un signal géopolitique. Avec des inquiétudes croissantes sur la souveraineté IA face à la Chine (cf. ",[42,1388,1390],{"href":1389},"\u002Ffr\u002Fnews\u002Fdeepseek-v4-modele-ia-open-weight-1000-milliards","DeepSeek V4 levant un tour à 1 000 milliards",[42,1392,1394],{"href":1393},"\u002Ffr\u002Fnews\u002Fdeepseek-tencent-alibaba-20-milliards-valorisation-premier-tour-avril-2026","Tencent\u002FAlibaba à 20 milliards","), les États-Unis sécurisent leur supply chain IA militaire en intégrant un ",[17,1397,1398],{},"outsider domestique"," (Reflection) à côté des big tech.",[65,1401,1403],{"id":1402},"les-implications-géopolitiques","Les implications géopolitiques",[13,1405,1406],{},"Le bouleversement va au-delà des États-Unis. Trois conséquences à suivre :",[321,1408,1410],{"id":1409},"_1-leurope-se-positionne","1. L'Europe se positionne",[13,1412,1413,1414,1418,1419,1423],{},"Avec ",[42,1415,1417],{"href":1416},"\u002Ffr\u002Fnews\u002Fuk-sovereign-ai-fund-500-millions-callosum-souverainete-britannique-avril-2026","le UK Sovereign AI Fund de 500 millions"," et la pression française pour ",[42,1420,1422],{"href":1421},"\u002Ffr\u002Fnews\u002Fcommission-enquete-gafam-souverainete-numerique-france-2026","une commission d'enquête sur la souveraineté numérique",", l'Europe accélère sa propre stratégie souveraineté. Mistral, AMI Labs (Yann LeCun), et probablement BAE Systems \u002F Thales ouvrent des programmes IA défense.",[321,1425,1427],{"id":1426},"_2-la-chine-et-la-russie-ré-arment","2. La Chine et la Russie ré-arment",[13,1429,1430,1431,1434],{},"L'exclusion d'Anthropic et la signature massive avec sept géants américains envoient un signal clair à Pékin et Moscou : ",[17,1432,1433],{},"les États-Unis intègrent l'IA dans leur chaîne militaire",". Réponse attendue : accélération du programme Tiangong de l'EPL et déploiement plus rapide de Yandex \u002F Sber Cloud sur les systèmes russes.",[321,1436,1438],{"id":1437},"_3-les-startups-défense-ia-explosent","3. Les startups défense IA explosent",[13,1440,1441,1442,1445,1446,1450],{},"L'inclusion de Reflection ouvre la voie à un ",[17,1443,1444],{},"deuxième cercle"," de startups défense IA. ",[42,1447,1449],{"href":1448},"\u002Ffr\u002Fnews\u002Fshield-ai-2-milliards-serie-g-aechelon-hivemind-defense-ai-avril-2026","Shield AI a déjà levé 2 milliards de Série G",", Anduril continue son expansion, et des nouveaux entrants (Scout AI, Helsing, Hadrian) lèvent des Séries A monstrueuses.",[65,1452,1454],{"id":1453},"ce-que-ça-change-pour-les-développeurs-ia-et-les-apps-grand-public","Ce que ça change pour les développeurs IA et les apps grand public",[13,1456,1457],{},"Pour les développeurs et apps grand public, l'impact direct est limité — mais le contexte global change :",[268,1459,1460,1466,1472],{},[271,1461,1462,1465],{},[17,1463,1464],{},"Les modèles publics restent identiques",". GPT-5.4, Claude Opus 4.7, Gemini 2.5 Pro continuent d'être disponibles via API standard",[271,1467,1468,1471],{},[17,1469,1470],{},"La concurrence technologique s'intensifie",". Avec des contrats DoD à 300-500 M$, OpenAI et Google ont un levier de financement supplémentaire pour pousser le frontier (GPT-6, Gemini 3.5)",[271,1473,1474,1477,1478,1482,1483,1486],{},[17,1475,1476],{},"L'arbitrage des talents bouge",". Comme on l'avait analysé dans le ",[42,1479,1481],{"href":1480},"\u002Ffr\u002Fnews\u002Fbrain-drain-meta-google-openai-startups-ia-188-milliards-vc-avril-2026","brain drain Meta-Google vers les startups",", une partie des chercheurs safety qui ne veulent pas travailler sur défense vont ",[17,1484,1485],{},"migrer vers Anthropic"," ou de nouvelles structures",[13,1488,1489,1490,1493,1494,1498],{},"Pour les éditeurs d'apps IA grand public, le message reste cohérent : la monétisation des produits IA passe par ",[17,1491,1492],{},"les utilisateurs civils, pas le DoD",". Le SDK chat et les formats publicitaires natifs comme ",[42,1495,1497],{"href":1496},"\u002Ffr\u002Fpublishers","ceux d'Idlen"," restent la voie la plus rapide pour transformer une app IA en produit rentable, sans dépendre des cycles de procurement gouvernementaux.",[65,1500,1502],{"id":1501},"conclusion-la-fracture-éthique-de-lia-américaine-est-consommée","Conclusion : la fracture éthique de l'IA américaine est consommée",[13,1504,1505,1506,1509,1510,1513,1514,1517],{},"Le 1er mai 2026 marque un ",[17,1507,1508],{},"point de bascule éthique"," dans l'industrie IA américaine. Pour la première fois depuis Project Maven en 2018, ",[17,1511,1512],{},"les sept plus grosses AI labs domestiques"," (Anthropic exclue) ont collectivement décidé que ",[17,1515,1516],{},"les contrats militaires classifiés valent l'entorse aux Usage Policies",". C'est un précédent qui va structurer les 10 prochaines années.",[13,1519,1520,1521,1524,1525,79],{},"Anthropic, en restant à l'écart, paie un prix business. Mais gagne un positionnement éthique qui pourrait, à terme, devenir un avantage commercial massif sur les segments santé, education, et clients européens sensibles à la souveraineté. Comme l'écrit Dario Amodei dans son dernier mémo interne : ",[17,1522,1523],{},"« Notre brand est notre policy. »"," Les prochains mois diront si cette discipline paie en termes de valorisation IPO — un sujet décisif pour ",[42,1526,1528],{"href":1527},"\u002Ffr\u002Fnews\u002Fanthropic-900-milliards-ipo-octobre-2026-50-milliards-tour-mai-2026","la trajectoire d'Anthropic vers une IPO en octobre 2026",[562,1530],{},[13,1532,1533],{},[567,1534,1535,1536,1540,1541,1544,1545,79],{},"Pour continuer sur les enjeux IA militaires et géopolitiques, voir notre coverage de ",[42,1537,1539],{"href":1538},"\u002Ffr\u002Fnews\u002Fpentagone-100000-agents-ia-vibe-coding-genai-mil-google-gemini-avril-2026","100 000 agents IA déployés au Pentagone",", notre analyse du ",[42,1542,1543],{"href":1091},"Trump-Anthropic ban initial"," et notre ",[42,1546,1548],{"href":1547},"\u002Ffr\u002Fblog\u002Fcomment-monetiser-app-ia-guide-complet","guide complet sur la monétisation des apps IA",{"title":584,"searchDepth":585,"depth":585,"links":1550},[1551,1552,1553,1554,1559,1560,1561,1566,1567],{"id":1062,"depth":585,"text":1063},{"id":1111,"depth":585,"text":1112},{"id":1231,"depth":585,"text":1232},{"id":1281,"depth":585,"text":1282,"children":1555},[1556,1557,1558],{"id":1288,"depth":593,"text":1289},{"id":1303,"depth":593,"text":1304},{"id":1324,"depth":593,"text":1325},{"id":1335,"depth":585,"text":1336},{"id":1367,"depth":585,"text":1368},{"id":1402,"depth":585,"text":1403,"children":1562},[1563,1564,1565],{"id":1409,"depth":593,"text":1410},{"id":1426,"depth":593,"text":1427},{"id":1437,"depth":593,"text":1438},{"id":1453,"depth":585,"text":1454},{"id":1501,"depth":585,"text":1502},"Le 1er mai 2026, le Pentagone a finalisé des accords IA pour les réseaux classifiés IL6\u002FIL7 avec huit géants tech, en excluant volontairement Anthropic. Le contrat de 200 M$ d'Anthropic est mort. Analyse complète des enjeux militaires, géopolitiques et éthiques.",{},"\u002Ffr\u002Fnews\u002Fpentagone-8-contrats-ia-classifies-openai-google-spacex-anthropic-exclu-mai-2026",{"title":1017,"description":1568},"Le Pentagone signe avec OpenAI, Google, Microsoft, AWS, Nvidia, SpaceX, Oracle et Reflection. Anthropic exclu après refus de modifier ses safety limits.","Pentagone 8 contrats IA classifiés mai 2026 : Anthropic exclu, OpenAI signé","fr\u002Fnews\u002Fpentagone-8-contrats-ia-classifies-openai-google-spacex-anthropic-exclu-mai-2026",[1576,619,620,621,622,623,624,625,626,627,628,629,630],"pentagone","5eb-oXJWg2FWg9H2ns6LvnbaXlG7ODgRwadPrM01564"]