Pentagon tests OpenAI and Google

The Pentagon has begun formal testing of artificial intelligence models from OpenAI and Google after restricting some uses of Anthropic’s Claude, according to Bloomberg and follow-on reports citing a senior defense official. The evaluation is aimed at finding alternatives for sensitive workloads inside the department, with 25 “power users” asked to compare competing systems. India Today, citing the Bloomberg report, said the review followed a breakdown in ties between the Defense Department and Anthropic over usage restrictions on Claude. The shift puts model choice, procurement flexibility and monitoring tools at the center of how high-security organizations are deploying generative AI. It also shows how quickly a large buyer can move to alternate suppliers when policy, safety or access terms change. ### Who is actually being tested, and by whom? Bloomberg reported on May 21 that the Defense Department is testing OpenAI and Google models to see which are preferred by 25 departmental “power users,” citing a senior defense official. India Today said the review is intended to replace Anthropic’s Claude in restricted contexts after the Pentagon stopped using some Anthropic systems. The Pentagon has not publicly detailed the exact model versions under review in the reporting now available. The department has, however, repeatedly said in broader strategy documents that it wants to design and test AI-enabled systems with stakeholders across operational environments and build the infrastructure, policies and models needed for department-wide adoption. ### What triggered the move away from Anthropic? India Today reported on February 24 that Defense Secretary Pete Hegseth had summoned Anthropic Chief Executive Dario Amodei as the department pushed leading AI developers to make advanced models available inside classified military networks with fewer restrictions. A later India Today report on March 13 said Pentagon officials had labeled Anthropic a supply-chain risk and that Anthropic was contesting that designation. (bloomberg.com) India Today and CXO DigitalPulse both said the dispute centered on Anthropic’s limits around certain military and surveillance-related applications. Those reports describe the latest testing of OpenAI and Google as part of the Pentagon’s search for substitute model providers after that clash over guardrails and access. ### Why does this matter beyond one vendor switch? (indiatoday.in) The Department of Defense’s own AI strategy documents say the department is building foundational enablers including infrastructure, data, models and policies, and that it intends to test AI systems across different environments before broader deployment. That makes vendor substitution a procurement and systems question, not only a product comparison. (indiatoday.in) For organizations deploying assistants on sensitive data, the Pentagon’s testing underscores a practical requirement: the model layer may need to be replaced without rebuilding the surrounding controls. The public reporting does not say the Pentagon used the phrase “vendor-agnostic observability,” but the move from one frontier model provider to others points to the value of keeping evaluation, logging and policy enforcement portable across vendors. That is an inference based on the reported testing process and the department’s published emphasis on repeatable evaluation and responsible AI controls. (media.defense.gov) ### What does the Pentagon say about AI controls more broadly? The Defense Department’s Responsible AI Strategy and Implementation Pathway says AI systems used by the military must be designed, developed, procured, deployed and used in ways consistent with lawful and ethical behavior. A separate 2026 strategy document says the department is using “pace-setting projects” to accelerate AI adoption while building the policy and infrastructure needed to support it. (bloomberg.com) Those documents do not mention this specific OpenAI-Google test, but they provide the framework in which the evaluation is taking place: rapid iteration, measurable outcomes and controls around deployment. ### What comes next in this review? Bloomberg’s report said the current phase is focused on feedback from 25 Pentagon power users, and the department has not yet publicly identified a winner or announced a procurement decision. (media.defense.gov) India Today said the testing is part of an effort to replace Claude in restricted settings rather than a full public overhaul of all Pentagon AI systems. (media.defense.gov) Any next step is likely to show up first in Pentagon contracting records, official defense statements or further reporting naming the systems selected for broader deployment. As of May 23, 2026, the public reporting points to an active evaluation of OpenAI and Google models, with Anthropic’s status still shaped by the earlier dispute over restrictions and risk designations. (bloomberg.com)