Pentagon replaces Anthropic in classified military AI contracts
Key insights
- The DoD finalized classified AI agreements with eight vendors on May 1, formally ending its $200M Anthropic contract after a March supply-chain-risk declaration.
- Early evaluations across 25 military power users show performance varies sharply with prompt phrasing, and no replacement model yet matches Anthropic's prior deployments.
- Defense Secretary Hegseth's supply-chain framing explicitly ties Anthropic's safety guardrails to its disqualification, not capability gaps.
Why this matters
The DoD's explicit labeling of safety guardrails as a supply-chain disqualifier gives every AI vendor selling into government a concrete signal to deprioritize alignment features in their federal offerings. The eight-vendor classified agreement structure puts OpenAI, Google, and newer entrants like Reflection AI in direct competition at the highest classification levels, accelerating capability races without safety constraints as a competitive factor. Prompt-sensitivity instability flagged in early military evaluations suggests the hardest remaining engineering problem is operational reliability under adversarial conditions, which will likely concentrate DoD contract flow toward vendors with the deepest fine-tuning and red-teaming infrastructure.
Summary
The Pentagon labeled Anthropic a supply-chain risk in March, citing its insistence on safety guardrails, and three days later launched rival evaluations across 25 military power users.
By May 1, the DoD had signed classified AI agreements for IL6/IL7 networks with eight vendors: OpenAI, Google, Microsoft, AWS, Nvidia, SpaceX, Oracle, and Reflection AI, formally replacing its $200M Anthropic contract. Defense Secretary Hegseth's framing made the stakes explicit: safety constraints are now officially a disqualifier in classified procurement.
Essentially: (OpenAI, Google, xAI) are racing to fill operational gaps Anthropic left, but none has matched its prior deployments yet.
- Model performance varies significantly with prompt phrasing, requiring ongoing optimization before classified deployments reach operational readiness.
- Reflection AI joins the classified vendor list alongside hyperscalers, a notable inclusion for a company of its age.
The supply-chain-risk designation sets a concrete precedent for how every AI vendor must now position its product to access classified government contracts.
Potential risks and opportunities
Risks
- Anthropic loses viability as a government vendor, potentially triggering renegotiations or cancellations across other federal contracts it holds beyond the $200M DoD deal
- Prompt-sensitivity instability documented in early DoD evaluations could cause model failures in high-stakes classified operations before optimization is complete, with no Anthropic fallback now in place
- The supply-chain-risk designation for safety guardrails creates a replicable template other federal agencies could use to pressure remaining safety-focused AI vendors out of procurement processes
Opportunities
- OpenAI and Google can leverage IL6/IL7 classified agreements as proof points to accelerate allied-nation government sales where top-tier security classification is a differentiator
- Reflection AI's inclusion in the classified vendor list positions it as a credible defense AI player, likely unlocking defense-focused investment and follow-on contract opportunities within the next 60 to 90 days
- Prompt optimization and red-teaming firms (Scale AI, Patronus AI) could see increased DoD contract flow as military users require specialized fine-tuning work to stabilize model performance under adversarial prompt conditions
What we don't know yet
- Whether any of the eight new vendors have committed to specific safety constraints for IL6/IL7 deployments, or if guardrail-free operation is now the implicit procurement standard
- How Reflection AI secured a classified AI agreement alongside hyperscalers given its relative youth, and whether it holds the necessary FedRAMP or ITAR compliance infrastructure
- Whether the $200M Anthropic contract termination will face legal challenge or congressional review before the end of Q2 2026
Originally reported by bloomberg.com
Read the original article →Original headline: Bloomberg: Pentagon Testing OpenAI and Google AI Models in Race to Replace Anthropic's Claude for Classified Military Systems