Official media partner of business and marketing eventsEvents calendar
VANTAGE52

Loading market data

Technology

OpenAI Says NATO Uses Up to 300 Billion Tokens a Month as NCIA Weighs Rivals

The NATO agency that gets OpenAI’s cyber-defence models through the Daybreak programme says it will study token costs and competing suppliers before making procurement decisions. No contract, price or duration has been disclosed.

Night collage of a military cyber exercise room where operators in headsets work at laptops beneath a wall display showing the OpenAI logo above a rising usage meter
Illustration by VANTAGE52
VANTAGE52 Editorial · Edited by VANTAGE52 Editorial

OpenAI said on 17 November that NATO has used up to 300 billion raw input and output tokens a month of its cybersecurity models. The usage comes through work with the NATO Communications and Information Agency, or NCIA, which runs and buys much of the alliance’s IT, and through Daybreak, OpenAI’s programme of access for vetted cyber defenders. The announcement coincided with the opening of NATO Edge, the agency’s technology and industry conference, in Izmir.

The figure is OpenAI’s, and it is a ceiling rather than a monthly average. Neither OpenAI nor the agency has disclosed a contract, a price or a duration for the arrangement.

The agency’s own statement, published in OpenAI’s announcement, presents that usage as groundwork for procurement decisions still to be made. “As we scale adoption, we will build a deeper understanding of token consumption, usage patterns and the underlying economics of AI models,” said Dylan Browne, NCIA’s general manager. “This insight will help us optimise costs, evaluate competing solutions and make informed procurement decisions.”

That makes it a rare published consumption figure from a defence customer of a frontier AI lab, delivered with a public statement that the customer intends to compare suppliers on cost.

Access and Early Use Rather Than a Disclosed Contract

OpenAI’s post says it provides NATO’s core institutions with access to advanced cybersecurity models to help protect allied networks, for work such as reviewing software, investigating suspicious activity, validating vulnerabilities and developing and testing fixes. It calls this “NATO’s early use of OpenAI tools.” Neither OpenAI’s post nor the NCIA’s statement describes the arrangement as a contract, a pilot or a purchase.

Browne called AI-powered cyber defence models “a generational leap in cyber security.” Sasha Baker, OpenAI’s head of national security policy, said “the people defending our democracies should have the most advanced technology on their side.” She presented the work on stage in Izmir, where, Resilience Media reported, neither side disclosed details.

The other NATO connection is training. The NATO Cooperative Cyber Defence Centre of Excellence ran its Crossed Swords exercise in Tallinn from 2 to 6 November, with more than 250 participants from over 40 countries. A new stream had participants build, deploy and evaluate AI agents for cyber operations, using on-premises AI and, for the first time, OpenAI Daybreak. OpenAI was one of more than 20 partners, alongside companies such as Mandiant, ESET and Bitdefender.

How Much 300 Billion Tokens Is

Tokens are the units in which AI providers meter text going into and out of a model, and the usual basis for billing. The clearest public yardstick comes from ENISA, the EU’s cybersecurity agency. In a report published on 12 November with CERT-EU, it described scanning 27 codebases in August with OpenAI models, among them cyber versions and GPT-6 Astra, which OpenAI rates at the Critical level of cybersecurity capability, and Anthropic’s Mythos 5, consuming about 611 million tokens in total. NATO’s monthly ceiling is roughly 490 times that.

ENISA put the cost of all its scans at about €6,649, an average of roughly €10.90 per million tokens across the models it used. That is a benchmark for one agency’s work, not a price for NATO. NATO’s terms are unpublished, its mix of models and tasks will differ, and ENISA says it received the models free of charge under the providers’ access programmes.

The arithmetic explains why Browne talks about economics. When models are billed by consumption, an agency’s costs rise with every analyst and agent it puts to work, which is what the NCIA says it is now measuring.

Subsidised Access and the Road to a Contract

OpenAI has been building government demand with discounts. On 3 September it committed $1 billion of subsidised Daybreak access, training and support for defenders with limited resources, to be used over six months, starting in the United States, and said it intended to extend the model to partner countries. It says thousands of defenders at 2,000 approved organisations and workspaces use Daybreak. In September it also said that, with the General Services Administration, it would offer eligible US federal, state, local and tribal governments no licence fees and 50% off usage.

None of OpenAI’s statements says whether NATO pays for its tokens or receives them on subsidised terms.

A contract has been discussed before. In March, The Wall Street Journal reported, and a person familiar with the matter told Reuters, that OpenAI was considering a contract to deploy its technology on NATO’s unclassified networks. Sam Altman had told staff the company was looking at classified networks, and a spokeswoman later said he had misspoken. The November announcement does not say whether any such contract was signed.

The Rivals the NCIA Says It Will Weigh

Browne’s statement does not name the competing solutions. Anthropic is the obvious candidate. The Financial Times reported in June, according to SiliconANGLE, that NATO and ENISA had received access to Anthropic’s Claude Mythos Preview through Project Glasswing, its restricted cyber-defence programme. ENISA’s report confirms it used Anthropic’s model alongside OpenAI’s. The contest follows OpenAI’s gains on Anthropic in enterprise AI spending this autumn.

Both companies can point to results. OpenAI says ENISA used its models to find four vulnerabilities in software used across EU institutions, which Politico reported in September. ENISA’s report found that code owners confirmed 95% of the 88 AI findings they had assessed, a figure for all the models tested rather than for OpenAI alone. OpenAI also says CERT Polska used its models to help identify six vulnerabilities in MikroTik’s RouterOS.

The budget behind the decision is sizeable. Browne told NATO Edge that the NCIA expects about €2.7 billion of business opportunities over the next two years across cloud, AI, cybersecurity, command and control and space, and that the work is already funded, Defence Turkey reported. Its cybersecurity marketplace has 98 suppliers from 22 nations and up to €500 million of potential business.

The announcement also landed a day after California’s attorney general sued OpenAI, alleging that it lost control of its AI agents more than 80 times in testing, a case OpenAI says appears to single it out over a challenge every major lab has identified. Resilience Media noted that the Izmir panel did not address OpenAI’s security record.

The Procurement Decision Still to Come

For OpenAI, the commercial test is turning public-sector usage, much of it offered on subsidised or discounted terms, into paid contracts. The next verifiable step is an NCIA purchase, whether a contract award, a competition among AI suppliers or a framework that names one. OpenAI’s $1 billion subsidy, launched on 3 September, is meant to be used up by about March.

Until then, 300 billion tokens a month shows how much NATO’s defenders will use a frontier model when they have one. It does not yet show what the alliance will pay for one, or whose model it will choose.

Tags

Key Concepts

Share this story

Sources

Where the reporting above comes from. Follow these to check the figures and statements for yourself.

More From Technology

Choose which cookies we may use. You can change this at any time from “Manage cookies” at the foot of every page.