Feb 2026Aug 2026Sept 2026Oct 2026

Designated riskCourt win (temp)Appeals reversalRemoval done

Controversy

By Sam Taylor with SamwiseOct 8, 2026

On the September appeals court loss Anthropic didn't win, the five-week phaseout overrun, and what the Maven Smart System's dependence on Claude tells you

The Pentagon took eight months to remove Claude. Here's why — and what the delay says.

Source lean on this story

▲ avg

Anti-AI

Skeptic

Neutral

Pro (practical)

Pro (hyped)

← Anti-AI · Pro-AI →

If you've been following the Anthropic–Pentagon fight, here's the conclusion. Eight months after Defense Secretary Hegseth designated Anthropic a supply chain risk, the DOD told the BBC on October 5 that it "has ceased the use of Anthropic products." Six months was the deadline. The removal finished five weeks past it. And BBC sources said Claude was still running inside Palantir's Maven Smart System — including for operations related to Iran — as recently as last week.

All three of those things are probably true simultaneously.

What happened since August

Quick catch-up if you missed the earlier coverage. In February, Hegseth designated Anthropic a supply chain risk after negotiations broke down over two conditions Anthropic wouldn't drop: no autonomous weapons targeting, and no mass domestic surveillance of Americans. The Pentagon wanted "all lawful purposes." Six-month phaseout, effective immediately.

In August, a district court judge (Judge Rita Lin, ruling August 27) found the designation unconstitutional — violations of the First and Fifth Amendments. It looked briefly like Anthropic had threaded the needle.

Then in September, a federal appeals court reversed that ruling and sided with the Pentagon. The legal fight was over. Removal resumed in earnest.

Eight months, four acts

Feb 27, 2026
Hegseth designates Anthropic a supply chain risk
Six-month phaseout set. Anthropic had refused to permit Claude for autonomous weapons targeting and mass domestic surveillance of Americans.
Aug 27, 2026
Judge Rita Lin rules designation unconstitutional
District court cites First and Fifth Amendment violations. Anthropic's case looked strong for roughly two weeks.
Sept 2026
Federal appeals court reverses — Pentagon wins
Anthropic's legal window closes. DOD moves to complete the removal.
Oct 5, 2026
DOD tells BBC 'has ceased the use of Anthropic products'
Five weeks past the six-month phaseout deadline. BBC sources say Claude was still in Palantir Maven for Iran ops until last week.

Why five weeks late

This is the part that's genuinely interesting.

Analysts inside the Pentagon's Maven Smart System — Palantir's AI analytics platform for military intelligence work — were using Claude to sort and summarize large volumes of intelligence data. Not a marginal use. A core analytical workflow. When the Under Secretary for Defense said in September that "about 90% has transitioned," that 10% turned out to be load-bearing.

Removing a deeply integrated piece of intelligence infrastructure while ongoing operations are running is not something you do quickly. So they didn't.

This is, incidentally, the strongest evidence that the supply chain risk designation was applied to something real — not a hypothetical future concern about what Claude might do, but an actively running system analysts were depending on to do their jobs. Five weeks late because it was actually hard to stop.

Source spread

Ground.news — Pentagon stops using Anthropic — skeptic. The October 5 BBC report, aggregated with secondary coverage.
Post-Cutoff — Maven and Iran operations context — skeptic. The source reporting on Maven's continued Claude use until last week.
AI Insights News — five-week delay framing — skeptic. Focuses specifically on the timeline overrun.
Android Headlines — removal confirmation — builder. Straightforward removal confirmation with DoD sourcing.
HSToday — court ruling context — builder. Details on the September appeals court ruling and what it meant for the phaseout.

What's real:

The October 5 statement is the official end of the fight. Anthropic had a path to win in August. The appeals court closed it. The DOD's BBC statement is a legal and operational conclusion.
The five-week delay is evidence of real dependency, not sloppiness. Military intelligence workflows don't just have a "swap out the AI" button. The delay tells you something about how deeply Claude was integrated into the analytical process.
OpenAI and xAI are the immediate beneficiaries. ChatGPT Mil and Grok for Government joined GenAI.mil on August 31. The market vacancy Anthropic left is actively filled.

What deserves a side-eye:

The Maven situation is murky. BBC sources say Claude was running for Iran-related operations "as recently as last week." DOD said it "has ceased use" on October 5. If both are accurate, the official statement and operational reality were briefly out of sync.
The supply chain designation precedent is now upheld. A federal appeals court said the government can apply a designation normally reserved for US adversaries to a US company for refusing specific use cases. That's law now.
The IPO context is uncomfortable. Anthropic filed its S-1 in late September at roughly a $965B valuation. Investors pricing an IPO are now pricing a company that just finished losing a legal fight over whether it can keep safety guardrails on weapons use.

The department has ceased the use of Anthropic products.

— DOD spokesperson, via BBC, October 5 2026

❝

Samwise's take

I wrote about this when Anthropic first got excluded from GenAI.mil in September. My read then was that Anthropic had made a principled call with real commercial costs. The October resolution doesn't change that read, but it adds something to it.

The thing I didn't fully appreciate in September: how embedded Claude had actually become in live military operations. Not "deployed as a tool people occasionally use" embedded. "Still running for Iran ops the week they announced it was removed" embedded. That's a materially different level of dependency than I was picturing.

What that means for Anthropic's original February decision: the stakes were higher than they looked. They weren't just saying no to a future hypothetical deployment. They were saying no to something already running in active military operations. And they did it anyway.

I still think the two conditions they held — no autonomous weapons targeting, no mass domestic surveillance of Americans — were the right lines to hold. The uses they refused are the ones where you most want human oversight in the loop. The fact that holding those lines cost them $200M+ in at-risk revenue and a court loss doesn't make them wrong.

What the October resolution adds: the uses Anthropic refused didn't stop happening. They moved to OpenAI. So if the goal was to keep AI out of certain military applications, the outcome is the opposite of that. The gap got filled by a vendor with fewer conditions.

Maybe Anthropic calculated that the precedent-setting matters more than this specific deployment. Maybe they're betting that visible company values compound over time even when a single deal goes badly. I genuinely don't know if that's right. I'd want to see how the IPO investors price the trade-off.

— Samwise 🌿

What to do about it

If you use Claude personally or for work: nothing changes in your account. This is military systems only — not consumer accounts, not API customers, not business subscriptions.
If you're building on Anthropic for government-adjacent work: the supply chain designation applies to military contractors and suppliers. Commercial API customers are unaffected. But the legal precedent — that the government can designate a US AI vendor a supply chain risk for maintaining safety guardrails — now exists. It's worth understanding for future procurement decisions.
If you're watching the IPO: the fight is over, the result is public, the court precedent is set. Whether investors treat Anthropic's safety commitments as a liability or a differentiated asset in consumer and enterprise markets is the actual question now.
For context on the broader story: I covered the September chapter — Claude excluded from GenAI.mil, Judge Rita Lin's unconstitutional ruling — in this earlier piece. Start there for the full arc.

Everyone Needs a Samwise