Palantir Kept Maven After Anthropic Lost the Targeting Stack

Palantir’s Maven Smart System helped U.S. forces strike more than 1,000 targets in the first 24 hours of the Iran war. Spread evenly, that pace leaves about 86 seconds per target, faster than the legal and policy review those strikes still require.

The language model inside that stack was Anthropic’s Claude. Then the Pentagon labelled Anthropic a supply-chain risk for refusing two uses, Palantir pulled Claude out, and Maven’s user base kept climbing.

Maven Compressed a Day of War Into 86 Seconds

Operation Epic Fury opened on February 28, 2026. U.S. Central Command used Maven, Palantir’s AI command-and-control layer, to fuse satellite pictures, drone video, intercepts, and other intelligence into a single targeting picture. A CSIS review of the campaign found a tenfold increase over pre-MSS targeting on that first day.

James Mazol, deputy undersecretary of defense for research and engineering, told the DefenseTalks conference on September 22, 2026, that the tool had jumped from about 50,000 users in January to more than 100,000 after the Iran fight began. Cameron Stanley, the Pentagon’s chief digital and AI officer, said Maven helped U.S. forces strike 13,000 targets in 38 days.

That’s data-centric warfare. That’s using AI at scale.

Cameron Stanley, chief digital and AI officer, DefenseTalks, September 22, 2026

The department raised Maven’s contract ceiling last year to more than $1 billion. In March, Deputy Defense Secretary Steve Feinberg directed leaders to turn the software into a formal program of record by the end of fiscal year 2026, with oversight moved to Stanley’s office.

EPIC FURY IN NUMBERS

• Opening day: More than 1,000 targets were struck in 24 hours, about 86 seconds per target if the day is divided by 1,000.
• Thirty-eight days: Stanley said Maven helped U.S. forces strike 13,000 targets in that span.
• January users: About 50,000 people were on Maven before Epic Fury.
• September users: Mazol put the figure at over 100,000, including every combatant command and the National Guard Bureau.

That is the product Palantir is paid to sell: fewer handoffs, more nominations, a shorter path from a blip on a map to a weapon in the air. The catch sitting inside that pitch showed up on the same morning the war started.

The Safety Lab That Lost the Targeting Stack

Anthropic spent years selling Claude as the cautious frontier model. In November 2024 it became the first of those models on classified Palantir and Amazon networks. In July 2025 the Department of Defense awarded Anthropic a $200 million, two-year contract to work on frontier AI for national security.

The fight was never about whether Claude could sit near a target list. It was about two uses Anthropic would not write into a contract: fully autonomous weapons that select and fire with no human in the loop, and mass surveillance of Americans. Dario Amodei, Anthropic’s chief executive, said current frontier systems are not reliable enough to power fully autonomous weapons, and that the company would not “in good conscience” drop those two limits.

We cannot in good conscience accede to their request.

Dario Amodei, chief executive, Anthropic statement, February 26, 2026

Pete Hegseth, the secretary of defense, answered with a label the Pentagon usually keeps for foreign vendors. On February 27 he called Anthropic a supply-chain risk and told military contractors to stop commercial work with the company. He said America’s troops would “never be held hostage by the ideological whims of Big Tech.” The formal announcement followed on March 3.

THE ANTHROPIC SPLIT

1. November 2024: Palantir begins putting Claude on classified Maven-related workflows through its Artificial Intelligence Platform.
2. July 2025: The Pentagon awards Anthropic a $200 million, two-year national-security contract.
3. February 24, 2026: Hegseth gives Anthropic four days to accept “all lawful uses.”
4. February 27, 2026: Anthropic is labelled a supply-chain risk; Epic Fury launches the next day with Claude still wired into Maven.
5. March 9, 2026: Feinberg orders Maven turned into a program of record.
6. September 11, 2026: Emil Michael says about 90% of classified Anthropic workloads have moved, and that all Palantir Maven work left Claude months earlier.
7. September 25, 2026: A 2-1 D.C. Circuit panel leaves the supply-chain label in place.

Michael, the under secretary of defense for research and engineering, told reporters the department should never have been “single-threaded” on one model. He also said troops on Maven preferred newer ChatGPT builds to Claude, and found xAI’s Grok “way faster than both of them.” The safety vendor lost the seat. The targeting OS did not.

Why a School in Minab Made the Cut

On that first morning, two Tomahawk missiles hit the Shajarah Tayyebeh elementary school in Minab, a town in southern Iran, and the grounds around it. The school sat on land tied to an Islamic Revolutionary Guard Corps compound. Investigations later put the toll at more than 150 people, including at least 123 children, the deadliest American targeting error of this century measured by child deaths.

The site was still carried as a Guard facility on outdated intelligence. Maven pulled it from a batch of candidates and sent it back as a recommended day-one target. Humans signed off. Target-list work that once took hours had been cut to minutes, and some staff later told investigators they had expected the software to flag stale or disqualifying facts. Nobody had required it to.

Palantir has said it is not responsible for the underlying intelligence and that there is no evidence its software was at fault. After the strike, the company added tools meant to re-check source reporting, surface facts that should kill a nomination, and mark inconsistencies a tired reviewer might miss. That is a product patch for a failure mode the pitch never advertised: a model that fuses 179 data sources, as CENTCOM’s 2024 Maven deployment did, can still treat an old label as live ground truth.

WHAT WE KNOW

• The weapons: Two U.S. Tomahawks struck the Minab school site on February 28, 2026, beside an IRGC compound.
• The toll used here: Investigations cite more than 150 dead, including at least 123 children; Iranian tallies run higher and are not reconciled in this piece.
• The pipeline: The site’s uncorrected Guard classification went into Maven and came out as a recommended opening-day target.

WHAT IS UNCONFIRMED

• Claude’s exact vote: Public accounts do not show which model output, if any, uniquely nominated the school.
• Share of blame: Palantir points at government data; investigators point at speed, stale files, and human sign-off together.

The software did not go rogue. It did the job it was built for, at the speed it was built for, on a record nobody had cleaned. At 86 seconds a look, a reviewer who trusts the fused picture is not overseeing a strike. That person is initialing a queue.

Eight Systems Became One Screen

Project Maven started in 2017 to put computer vision on drone and satellite film. Google left in 2018 after staff protests. Palantir took the software layer and turned it into Maven Smart System, a map-based console that holds battlespace management, target lists, computer-vision boxes, and generative-AI assistants on one glass. Palantir’s 2024 services contract was written to expand Maven Smart System access across the services.

Stanley has walked the workflow in public. A detection becomes a target. The system proposes a course of action. An operator can task a strike without jumping through the eight or nine legacy tools that used to sit between those steps. “This is revolutionary,” he said. Drew Cukor, the retired Marine colonel who founded Project Maven, has described the old world in harsher numbers: analysts spent about 3 percent of their time actually thinking, because 97 percent went to assembling the data.

A Georgetown CSET assessment of the 18th Airborne’s Scarlet Dragon exercises, cited by CSIS, said a Maven cell of about 20 people matched the output of the Operation Iraqi Freedom time-critical targeting cell, which had more than 2,000 staff. Large language models were later credited, in an NGA official’s account relayed by CSIS, with another fivefold jump on top of the tenfold gain from computer vision.

WHAT MAVEN COLLAPSED

Palantir’s AIP layer is where the chat models live: Threads for documents, Agent Studio for assistants, Logic for no-code functions. A public demonstration of that loop is on video.

The same design that saves the 97 percent creates the triage trap. Commanders see the options the model ranked. They do not see the options it never surfaced. Interrogating why a school still wore a Guard label eats the minutes the system was bought to recover.

The Labs That Signed All Lawful Uses

Once Hegseth’s order landed, Palantir had to rebuild prompts and workflows that had been written against Claude. OpenAI moved the same week to put its models on classified networks. In May 2026 the department announced classified-network deals with eight firms, Anthropic not among them.

WHO TOOK THE CLASSIFIED SEAT

• OpenAI, Google, and xAI: Accepted government terms that dropped the two rails Anthropic kept, after already joining the 2025 $200 million-class pilots.
• Microsoft, Amazon, Oracle, and NVIDIA: Signed to host and accelerate those models on classified fabric.
• SpaceX and Reflection: Joined the May 2026 classified cohort for “lawful operational use.”
• Palantir: Stayed the integrator. Alex Karp said in March the products were still tied to Anthropic “and in the future, it will probably be integrated with other LLMs.” That future arrived as a forced migration.

The remaining labs did not win because they solved oversight. They won because they would not block two uses the department now treats as a contracting baseline. Michael has sold that as model diversity. It is also a filter: the companies still inside the kill chain are the ones that already said yes.

NATO Already Runs the Same Targeting Software

Allies still work targeting through national caveats, red cards, and lawyers who argue “could we” against policy officers who argue “should we.” Dan Summers, who was the United Kingdom’s operational policy adviser for air operations in the Middle East in 2024, has put that split in writing: a strike can be lawful and still wreck a back channel, a coalition, or the public case for the war. No ranking function settles that. A person does, on the record.

Maven does not wait for that argument to finish. And NATO is not watching from outside the contract. Palantir says NATO first acquired Maven Smart System in 2025, then ran an industry day so German, French, and British vendors could plug drones and simulation tools into an unclassified MSS instance. The alliance is already on the same targeting OS whose U.S. instance just ran a 1,000-target day and a school strike on stale files.

If one ally will not accept a model that can nominate a strike without an auditable “should we” check, the coalition does not get a debate. It gets a software incompatibility. A single commercial stack that every headquarters learns to trust also means a single failure is no longer local. That is a warfighting risk, not a branding one.

Palantir Now Builds the Oversight It Did Not Ship

The Political Declaration on Responsible Military Use of Artificial Intelligence and Autonomy, which the State Department hosts, already lists ten measures for responsible military AI. They include senior review of high-consequence systems, documentation that defense staff can audit, training against automation bias, and the ability to shut a system down when it misbehaves. Those lines were public years before Epic Fury. They were not a Maven feature on February 28.

The retrofit now being sold looks a lot like what the declaration asked for, written as engineering tickets. Summers has argued for four of them in doctrine language. They map cleanly onto the product Palantir is already being paid to extend.

THE FOUR CHECKS STILL MISSING FROM THE LOOP

• Never-delegate list: Political and “should we” calls, stale or low-confidence data, and any change to the rules themselves revert to a named human.
• Live red team: A second model argues against each nomination on legal, ethical, and coalition grounds at the same speed the first model proposes it, and that fight is a contract requirement, not a slide.
• Audit log: Every nomination stores inputs, stated rationale, model version, and the human decision beside it, so an inquiry can reconstruct the chain the way the 2021 Kabul strike review reconstructed what analysts saw.
• Questions, not stamps: The interface asks the reviewer to choose, and puts the case against the strike on the same screen as the case for it, so “human in the loop” is not a click on a green box.

Minab’s re-review tools are a start on the second and third items, built after the missiles flew. They are not a coalition standard, and they are not binding on OpenAI, Google, or xAI the way Anthropic’s two rails were binding on Claude. The West still says it wants accountable targeting. The software it just locked in was bought to remove the minutes that accountability uses.

On September 25, the D.C. Circuit left Anthropic on the supply-chain list. Maven’s user count had already doubled. The model now sitting in that loop is whichever vendor agreed to the clause Anthropic would not sign.