OpenAI Reaches Pentagon Agreement While Anthropic Faces Exclusion? Verification and Implications (As of March 4, 2026)

Executive Summary

• The OpenAI-DoW relationship is not a single fresh award. It appears as a layered structure: a June 16, 2025 CDAO prototype OTA (up to 200 million USD) plus a later classified-environment agreement announced in late February 2026.
• The Anthropic exclusion narrative is less about losing a bid and more about three linked actions: a federal stop-use direction, GSA platform removal, and practical downstream restrictions in the defense supply chain.
• OpenAI publicly disclosed red-line language on domestic surveillance and autonomous weapons command, but the full contract text and complete procurement metadata are still not publicly available as of March 4, 2026.
• The core policy conflict appears to be between broad lawful-use demands by government stakeholders and narrower contractual safeguards demanded by AI providers.
• The highest-stakes open questions are legal validity of supply-chain-risk actions, auditability inside classified AI operations, and the gap between public clauses and real-world implementation.

What Is Confirmed About Contracts, Dates, Offices, and Value

Naming and Institutional Context

Public materials and reporting after February 2026 commonly use the label Department of War, often mapped to the same institutional context previously labeled DoD. This article treats Pentagon, DoW, and the relevant defense authority as the same context for timeline analysis.

Confirmed 2025 CDAO OTA for OpenAI

War.gov contract releases show that on June 16, 2025, CDAO awarded OpenAI Public Sector LLC a prototype OTA (HQ0883-25-9-0012) with a ceiling of 200 million USD. The stated purpose was frontier AI prototyping for war-fighting and enterprise missions, with performance centered in the National Capital Region and expected completion around July 2026.

Late-February 2026 Classified Agreement: Existence Confirmed, Procurement Metadata Not Fully Public

OpenAI stated on February 28, 2026 that it had reached an agreement with the Pentagon to deploy advanced AI systems into classified environments. Reuters also reported the existence of an agreement for deployment on classified defense cloud networks.

However, daily contract disclosures above the public threshold on War.gov did not show a clearly matching new OpenAI line item in the immediate Feb 27, Mar 2, and Mar 3 publication windows. As of March 4, 2026, a strict reading leaves two plausible structures:

• A modification or extension under pre-existing vehicles such as the 2025 OTA.
• A separate arrangement whose amount, format, or reporting status is not publicly disclosed.

So the key procurement fields for that late-February classified agreement remain unconfirmed: contract number, total value, term, and contracting authority.

Main Clauses OpenAI Publicly Shared

OpenAI disclosed relevant language excerpts, not a full contract. The disclosed points include:

• Lawful-use framing: use for all lawful purposes when aligned with applicable law, operational requirements, and safety/oversight protocols.
• Autonomous weapons boundary: no autonomous command role where human control is required, with explicit reference to DoD Directive 3000.09 style V and V and T and E controls.
• Surveillance boundary: no unconstrained monitoring of private information of U.S. persons; references to constitutional and statutory frameworks.
• Domestic law enforcement boundary: no use beyond legal limits such as Posse Comitatus constraints.
• Implementation controls: cloud-only architecture, managed safety stack, and cleared OpenAI personnel involvement.

OpenAI also reported a March 2 update clarifying that domestic surveillance use is prohibited, including cases involving commercially acquired personal or identifiable information, and that intelligence-agency use would require separate agreement.

Publicly Known Value and Duration

• OpenAI OTA: up to 200 million USD, expected completion July 2026.
• Anthropic OTA: up to 200 million USD, expected completion July 2026.
• OpenAI classified agreement announced in February 2026: amount and duration not publicly confirmed.

Reconstructed Timeline

• 2025-06-16: CDAO awards OpenAI OTA (HQ0883-25-9-0012), ceiling 200 million USD.
• 2025-07-14: CDAO awards Anthropic OTA (HQ0883-25-9-0014), ceiling 200 million USD.
• 2026-02-09: Non-classified GenAI.mil and ChatGPT integration is announced.
• 2026-02-26 to 2026-02-27: Anthropic and government-facing conflict escalates publicly.
• 2026-02-27 to 2026-02-28: OpenAI announces classified-environment agreement and publishes clause excerpts.
• 2026-03-02: OpenAI reports additional clarifying language on domestic-surveillance restrictions.
• 2026-03-04: Reuters reports defense contractors moving to remove Anthropic tooling from relevant workflows.

Why the Anthropic Exclusion Story Matters

The exclusion narrative is better explained by administrative and procurement mechanics than by normal competition outcomes.

First, government-side direction reportedly pushed Anthropic technology out of federal usage pathways. Second, GSA publicly announced exclusion from USAi.gov and MAS channels, which can function as institutional access denial even before final litigation. Third, defense-prime and supplier behavior appears to be adapting quickly, creating practical exclusion effects even where legal foundations are still contested.

Anthropic statements frame the dispute around two non-negotiable exceptions: mass domestic surveillance and fully autonomous weapons. The company also signaled planned legal action over supply-chain-risk designation and argued that legal scope should be limited rather than government-wide.

Legal and Policy Context

Autonomous Weapons Governance

DoD Directive 3000.09 is central because it emphasizes human judgment, validation and verification, and realistic testing and evaluation for autonomous or semi-autonomous systems. OpenAI's published language appears designed to align with this policy line.

Surveillance, EO 12333, and CAI Risk

The surveillance question is not only legality under existing authorities. The harder issue is whether large-scale aggregation and inference across commercially acquired and publicly available data can recreate practical surveillance effects even when formal red-line language exists.

Procurement and Supply Chain Risk Authority

10 USC 3252 is often cited in this dispute because it allows supply-chain-risk mitigation actions under specific conditions, typically tied to formal findings and oversight procedures. The unresolved question is whether required procedural steps were fully executed and how broad any resulting exclusion can legally be.

Impact Assessment

OpenAI

• Short term: strategic gain from classified-access agreement, paired with immediate scrutiny and clause-adjustment pressure.
• Medium term: external pressure to prove enforceability of safeguards in actual operations, not only in public summaries.
• Long term: credibility depends on whether cross-industry governance working groups produce auditable standards rather than high-level principles.

Anthropic

• Commercial access risk: removal from federal acquisition channels narrows entry points.
• Supply-chain spillover: defense contractors may avoid tools viewed as compliance or procurement risk.
• Litigation centrality: court outcomes could define precedent for emergency-style procurement and AI governance powers.

Industry and Strategic Competition

This episode may establish a broader precedent that lawful-use framing by the state can override narrower vendor guardrails in high-security contexts. That can reward vendors best aligned with government procurement expectations, while simultaneously increasing trust risk with parts of the technical labor market and civil-liberties stakeholders.

Key Unknowns as of March 4, 2026

• Full procurement metadata for OpenAI's classified agreement: contract number, amount, term, and legal vehicle.
• Documentary proof of procedural completion for supply-chain-risk actions under cited statutes.
• Official legal scope of exclusion: DoW-only contract context versus wider federal application.
• Operational interpretation of surveillance guardrails for CAI and large-scale public-data inference.
• Litigation trajectory, injunction outcomes, and potential congressional oversight steps.

Conclusion

A simple winner-loser framing is inaccurate. Available evidence supports a multi-layer event: existing OTA frameworks, subsequent classified agreement language, administrative exclusion mechanisms, and rapid supply-chain adaptation.

The decisive policy question is not who won headlines in one week. It is whether legal process, contractual safeguards, and operational auditing can be made simultaneously enforceable inside classified AI deployment environments.