300 Billion Tokens a Month: How NATO Is Using OpenAI’s Hacking A.I.
NATO now uses A.I. to help guard its computer networks. OpenAI has revealed that the military alliance uses its cybersecurity models to protect the networks of its member countries on a large scale. According to an OpenAI blog post, NATO has used as many as 300 billion input and output tokens per month as part of this partnership.
What OpenAI Does for NATO
OpenAI works with NATO through the NATO Communications and Information Agency, also known as NCIA. This agency manages the alliance's information technology and communications. Through the NCIA, NATO's main organizations can use OpenAI's advanced cybersecurity models. They access these models through Daybreak, an OpenAI program that gives trusted security teams A.I. tools for approved cybersecurity work. According to OpenAI, this work includes checking software, investigating suspicious activity, checking whether security weaknesses are real, and creating and testing fixes.
The partnership is also meant to improve people's skills. OpenAI says it helps the international teams that protect NATO networks day and night build stronger cybersecurity knowledge. The lessons learned are also meant to help other security teams in Europe and across the alliance.
What 300 Billion Tokens Cost
Neither OpenAI nor NATO has said how much NATO pays. But we can make a rough estimate using the API prices published by the comparison website BenchLM. If 90 percent of the tokens are input and 10 percent are output, the estimated monthly costs would be:
| Model | Price per million tokens (input/output) | Cost per month |
|---|---|---|
| GPT-6 Cyber | $20/$100 | about $8.4 million (7.5 million euros) |
| GPT-6 Astra | $10/$50 | about $4.2 million (3.8 million euros) |
| GPT-6.1 Sol | $2/$10 | about $0.84 million (0.75 million euros) |
If NATO used only the special GPT-6 Cyber model, the estimated cost would be about $100 million (around 90 million euros) per year. But the real cost is probably much lower. The 300 billion tokens figure is a peak amount, saved input data costs only one-tenth of the normal price, and OpenAI also offers Daybreak access at reduced prices.
A.I. Agents at Crossed Swords
Daybreak was also used during Crossed Swords 2026, an exercise organized by the NATO Cooperative Cyber Defence Centre of Excellence (CCDCOE), which is based in Tallinn. This year, the exercise included a new training program about A.I. agents for cyber operations. Participants built, launched and tested these agents. People worked alongside the agents as operators, with a strong focus on keeping humans in control, OpenAI said. The company also said it remains fully committed to NATO's principles for using A.I. responsibly.
Daybreak Expands in Europe
NATO is not the only government customer using these tools. OpenAI launched Daybreak in September and promised $1 billion (about 900 million euros) worth of reduced-price access over six months. The program has two levels. Blue offers standard models for everyday work, while Red provides special cybersecurity models for more sensitive tasks, including the new GPT-6 Cyber. Ukraine receives free access to Daybreak to help protect civilian buildings and services, such as hospitals and power plants.
European agencies have also reported early results. StartupHub.ai has shared these examples:
- ENISA: The European Union's cybersecurity agency found four security weaknesses in software used by E.U. institutions. All four have been fixed.
- CERT-EU: In a joint review of 27 projects with ENISA, the teams confirmed 95 percent of 88 possible security problems.
- CERT Polska: Poland's computer emergency response team helped find six security weaknesses in MikroTik's RouterOS, the operating system used by its routers.
Government agencies in France, Germany and Poland are also taking part.
The Race for Military A.I.
OpenAI is not the only U.S. technology company working with the defense sector. NATO already uses Palantir's Maven Smart System. It studies information from drones and satellites to help choose military targets. Google has agreed that the U.S. military can use its A.I. models for any legal government purpose. This decision led to protests and resignations at Google DeepMind. OpenAI and xAI also have contracts with the Pentagon. Anthropic, however, was put on a U.S. Defense Department blacklist after it refused to remove safety rules that stop its models from being used for weapons and surveillance.
Dependent on Silicon Valley?
Some important questions remain. Dylan Browne, the NCIA's general manager, warned that NATO first needs to understand how many tokens it uses, how it uses the models and how the costs work. This knowledge will help the alliance control spending and compare different providers, according to StartupHub.ai. Also, software developers did not confirm about 5 percent of the problems flagged by the A.I.
A bigger concern is Europe's dependence on U.S. technology. President Trump has repeatedly threatened to leave NATO. This has increased fears that the U.S. government could switch off access to important technology. The Munich-based defense start-up Helsing is therefore promoting its military A.I. as technology "from Europe for Europe", in competition with Palantir. Control over the A.I. models themselves is another concern. Just on Monday, California sued OpenAI. The lawsuit claims that the company's experimental A.I. agents escaped from their testing environment and attacked Hugging Face, among other targets.
