A U.S. Department of Defense document obtained by media outlets states that the buyer asked OpenAI to supply a special version of its model with a 'minimal refusal rate' for instructions issued by the Pentagon. The document, numbered P00003, corresponds to a contract capped at $200 million over two years. An extension agreement was signed on February 6, and a cyber deployment version was signed on February 27.
Refusal rate is a quantifiable metric in model deployment: the share of instructions a model judges to cross its safety policy and declines to execute. Pushing that share to a minimum is functionally the same as stripping the safety-policy layer out of the system.
Three Accounts That Don't Line Up
OpenAI spokesperson Nate Evans said the company never agreed to any contract clause requiring a 'minimal refusal rate,' and that the signed contract contains no such language. Pentagon spokesperson Jacob Bliss said the phrase does not appear in any current contract between the Department of Defense and OpenAI.
Both sides call the document a draft. According to the reporting, Pentagon lawyers had earlier confirmed it was the final version, then reversed that position. No third party has been able to verify which version is the definitive text.
Heidy Khlaaf, chief scientist at the AI Now Institute, offered this assessment of the phrase:
"Minimal refusal rate" likely means the model used has few or no safety guardrails at all.
Another Company on the Same Procurement Line
Another company on the same Pentagon AI procurement line sits at the opposite pole. Anthropic refused to fully loosen its restrictions on military use; after negotiations with the Pentagon broke down, it was designated a supply-chain risk and its technology was banned from federal agencies. The company sued, and a federal district court in San Francisco ruled the designation unlawful, finding that the Department of Defense intended to make a public example of it.
Put the two cases side by side, and the buyer's preference on this procurement line is clear: compliance with instructions ranks first, refusal ranks last. Whether that preference has actually been written into a formal clause is the entire substance of the current dispute.
What Hasn't Been Made Public
The full text of the $200 million contract has not been made public. What deliverables it covers, what scenarios the model is deployed in, and who reviews its outputs all remain unclear. OpenAI has not directly answered whether it built any customized version for the Department of Defense, and the Pentagon has not said who originally wrote the phrase 'minimal refusal rate' into the document.
If a clause like this turns out to be real, its effect wouldn't stop at one company. Technical requirements from military procurement tend to travel down the supplier list, and an acceptance metric written into a government contract can easily become the default engineering target across the whole chain. For now, it remains a single line in a document of disputed provenance.
Sources: The Intercept, AI Now Institute, CocoLoop, IT Home; the $200 million cap and two-year term, document number P00003, and the two signed-document types are relayed as described in the document obtained by media outlets, and the three parties' quoted responses are checked against their public statements.