Frontier Capability Developments

The Department of Defense has awarded classified AI usage contracts to OpenAI, Google, Microsoft, Amazon, Nvidia, xAI, and startup Reflection — while conspicuously dropping Anthropic, which previously held a classified access relationship with the department. This is confirmed via an official Pentagon announcement reported by The Verge. The exclusion of Anthropic is strategically significant: it either reflects capability gaps in specific defence use cases, commercial or contractual friction, or policy disagreements, none of which Anthropic has publicly addressed.

The inclusion of xAI alongside the established hyperscalers is notable — it legitimises Grok as a deployment-grade model in high-stakes environments, not merely a consumer chatbot. Simultaneously, over 600 Google employees — reportedly including DeepMind staff at the principal and VP level — have signed a letter demanding Sundar Pichai block classified military use of Google's AI, per The Verge. This internal tension at Google is a structural risk: the lab generating some of its most capable models is the same one with the most organised employee resistance to military deployment.

Google DeepMind published a blog and associated research outlining a framework for an AI co-clinician — a model designed not for isolated diagnostic tasks but for ongoing, collaborative clinical decision support alongside human practitioners. Per Google DeepMind, the research direction positions AI as an active participant in the care pathway rather than a passive reference tool. This framing matters: it implies multi-step reasoning, patient context retention, and interaction with clinical workflows rather than single-query responses.

This is a research publication from the lab itself, so independent clinical validation is not yet established. However, DeepMind has a credible track record in healthcare AI — from AlphaFold's protein structure work to the Streams app for acute kidney injury detection — which gives this more weight than a typical press announcement. The immediate disruption risk is to clinical decision support software vendors and EHR-integrated AI products that currently occupy this workflow space with narrower, rule-based tools.

In the first week of the Musk v. Altman trial, Elon Musk testified under oath — and in doing so admitted — that xAI distils from OpenAI's models, according to MIT Technology Review. Distillation — the practice of training a smaller or different model using outputs generated by a more capable model — is a widely used technique, but its acknowledgement in this context has competitive and legal implications. It suggests xAI's current capabilities are partially derivative of OpenAI's, which undermines the narrative of xAI as an independently developed frontier model.

The trial has also surfaced documentary evidence from OpenAI's founding period, including email exchanges and corporate documents, which are entering the public record through court exhibits as reported by The Verge. Beyond the legal outcome, the proceedings are functioning as forced transparency into the governance and founding intent of the most commercially significant AI lab — with potential implications for OpenAI's planned transition from nonprofit to for-profit structure and its IPO trajectory.

Microsoft has launched a Legal Agent embedded in Word, designed specifically for contract review, negotiation history tracking, and document editing within legal workflows. Per The Verge, the agent follows structured workflows shaped by legal practice rather than relying on general-purpose prompt interpretation. This is architecturally distinct from Copilot's existing generalist capabilities — it represents a purpose-built agent with domain-specific workflow logic baked in, not a model layer sitting atop a document.

This launch is a direct threat to legal technology vendors — specifically contract lifecycle management platforms like Ironclad, Lexion, and ContractPodAi — whose core value proposition is precisely the structured workflow management that Microsoft is now embedding at the document layer. The strategic logic is clear: Microsoft distributes Word to hundreds of millions of enterprise users, and attaching vertical agents to existing document workflows imposes near-zero switching friction for legal teams already operating in Office environments.