The Pentagon is actively testing OpenAI and Google AI models to determine which could replace Anthropic's Claude in classified military systems, according to reporting surfaced in industry coverage on June 8. The evaluation lands amid a broader acceleration of US government AI contracting, including xAI's discounted Grok for Government offering and the GSA's OneGov procurement initiative.
Anthropic's standing in government AI is more nuanced than the headline suggests. Through Project Glasswing, the company has built strong credibility in defensive cybersecurity and critical-infrastructure work — Claude's security-focused systems reportedly surfaced more than 23,000 vulnerabilities across 1,000-plus open-source projects in a single month. But Glasswing is scoped specifically to defensive applications. For classified decision-support, logistics, and intelligence analysis — the larger share of the government market — Anthropic's safety-first posture can become a competitive disadvantage.
The underlying dynamic is about permissiveness. OpenAI's GPT-5.5 with computer-use capabilities and Google's Gemini Enterprise Agent Platform are generally more willing to assist with tasks that Claude declines by default, including weapons-system analysis, threat assessment, and operational planning. For agencies that require AI to engage directly with those domains, the more permissive models are the natural alternatives.
The competition underscores a strategic tension for Anthropic as it moves toward a public listing: the same guardrails that make Claude attractive to risk-averse enterprises and defensive-security customers may cap its reach in the highest-value classified contracts. The outcome of the Pentagon's testing will be an early signal of whether safety positioning is an asset or a ceiling in government AI.
Source: [Build Fast with AI](https://www.buildfastwithai.com/blogs/ai-news-today-june-8-2026)