Pentagon Set to Finish Moving Classified AI Workloads Off Anthropic by October
The Pentagon expects to finish removing Anthropic's models from its classified AI workloads by the end of October 2026, with roughly 90% of that migration already complete. The remaining work includes systems tied to Palantir's Maven Smart Systems, the data-fusion platform at the center of the department's AI programs. Once the last workload moves, no frontier-model vendor predating the current procurement cycle will remain inside the US military's classified stack.
The shift ends a relationship that once made Anthropic the only frontier lab operating inside US military classified networks. It also turns a policy argument about usage limits into a procurement result. The department is buying model access it can direct without contractual conditions, and it is willing to rebuild its tooling to get it.
A Dispute About Contract Terms, Not Capability
The break did not follow a benchmark failure. Negotiations collapsed over Anthropic's usage restrictions, which the Department of Defense argued constrained lawful military applications. Defense Secretary Pete Hegseth delivered an ultimatum earlier this year, and the department went on to designate Anthropic a national security supply chain risk, warning that adversaries could exploit vulnerabilities in the technology and put warfighters at risk.
Anthropic has said it will challenge that designation in court, and the dispute is now before the D.C. Circuit. The White House separately ordered federal agencies to stop using Claude, while giving the Pentagon six months to phase the product out of systems where it was already deeply embedded.
That sequence explains why the exit is being run as a supply chain action rather than a routine contract termination. The designation obliges other companies with defense business to stop using Anthropic products within 180 days, which spreads the effect well beyond the department's own licenses. A directive of that shape reaches every prime contractor that has to justify its supplier list.
What Replaces Anthropic in Classified AI Workloads
Military users now work across a multi-model stack. OpenAI's ChatGPT and xAI's Grok carry most of the load, and the department reports that Grok has shown higher speeds while ChatGPT was preferred for certain mission attributes over the Anthropic models it replaced.
The Pentagon's chief digital and AI officer, Cameron Stanley, has said engineering work is underway to deploy alternative large language models inside government-controlled environments, and that completing the transition could take more than a month even after the bulk of the migration is done.
The replacement bench is wide. In May 2026 the department announced agreements with seven AI companies to deploy their systems on classified networks, a list that pointedly excluded Anthropic. Several of those relationships rest on contracts signed in July 2025, each carrying a ceiling of up to $200 million.
| Provider | Contract ceiling | Role in classified AI |
|---|---|---|
| OpenAI | Up to $200M | ChatGPT; preferred for certain mission attributes |
| xAI | Up to $200M | Grok; higher speeds in department testing |
| Up to $200M | Among seven firms deploying on classified networks | |
| Anthropic | Up to $200M | Designated supply chain risk; being phased out |
Contract ceilings describe the maximum the department may spend, not what it has spent. The structure of the agreements still carries a signal: several frontier labs signed comparable deals, and the one that attached conditions to how its models could be used became the one the department decided it could not route sensitive work through.
What the Multi-Model Stack Buys, and What It Costs
The department's answer to vendor risk is redundancy. Spreading classified work across ChatGPT, Grok and other providers means no single lab's policy decisions can halt operations, which is the exact exposure the Anthropic dispute created. The trade-off is integration overhead: multiple model families mean multiple sets of guardrails, evaluation pipelines and update cycles for teams that previously maintained one.
Iteration speed carries a second cost. When a single provider ships a new model version, one integration path absorbs the change. With several vendors inside classified environments, every update has to clear its own accreditation before it reaches users, which slows the refresh cycle the department once got from a single supplier.
The Commercial Bill for Anthropic
For Anthropic, the loss works on two levels. It removes a marquee federal deployment and the company's remaining footprint in classified workflows at the moment it is preparing a public listing. It also puts a supply chain designation on the record that defense contractors must weigh when they choose vendors.
The second effect lasts longer. A cancelled contract is a line item; a designation follows a company into every future procurement where a prime contractor has to defend its supplier choices. Anthropic's stated position is that the designation is unjustified and should be overturned, which is why the litigation in the D.C. Circuit carries commercial weight alongside its legal stakes.
OpenAI and xAI gain something harder to buy than revenue. Operating on classified networks at scale gives both firms validated defense references, the credential that shortens the next sales cycle inside the defense market. Google holds a comparable contract and sits in the same position.
The Dual-Use Fault Line Underneath
Emil Michael, the Pentagon's chief technology officer, has rejected warnings from an Anthropic researcher about catastrophic AI risk, describing them as a fear of change rather than a forecast the department should plan around. The exchange captures the gap between the two organizations. One treats limits on model behavior as a safety commitment; the other treats them as constraints on lawful military use.
The underlying problem is dual use. A model built for a broad commercial market and later adapted for classified systems carries assumptions that were never written for military parameters, and the department has now concluded it prefers providers whose terms are drafted for the defense mission from the outset.
That preference has its own price. Rewriting prompts, evaluation harnesses and integration code around new models takes engineering time, and the department's schedule has already stretched once. The 180-day clock set for contractors runs out this month, while the final military workloads are now expected to follow by the end of October.
The July 2025 contracts set uniform ceilings across providers. The department bought optionality rather than commitment, which is why the Anthropic exit required no renegotiation of the agreements themselves. The switching cost sat almost entirely in integration work and accreditation rather than in the paperwork.
What to Watch
The clearest near-term marker is the October timeline itself. A second slip would suggest the integration burden of a multi-vendor stack is heavier than the procurement decision implied. The D.C. Circuit's handling of Anthropic's challenge is the other variable: a reversal would not automatically restore classified access, but it would remove the supplier stigma that currently follows the company into contracting conversations.
The wider question is whether the Pentagon's approach becomes a template. Treating a vendor's usage conditions as a supply chain risk, rather than a commercial disagreement, gives government buyers a tool that reaches past the contract itself and into the vendor's entire defense-adjacent customer base. Any lab weighing similar restrictions now has a worked example of what that decision costs.
Why this matters
In defense procurement, control over how a model may be used has outweighed model quality. That trade is now visible to every enterprise buyer negotiating AI terms, governments included. For Anthropic, the coming months decide whether one designation becomes a lasting handicap or a reversible dispute as the company approaches a public listing.
Related Articles
- Pentagon AI Contracts Awarded to Seven Tech Giants as Anthropic Faces Exclusion
- Anthropic Mythos 5 Clearance Paves Way for Critical Infrastructure AI Deployment
- Court Voids Pentagon's Anthropic Blacklist as Retaliation
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.