Blog
NATO Is Using OpenAI’s Daybreak Cyber AI at Massive Scale: What 300 Billion Tokens a Month Really Means
17 November 2026
NATO is using OpenAI’s Daybreak program, a set of cybersecurity-focused AI models, to help protect Allied networks. According to OpenAI, usage has reached up to 300 billion tokens in a single month. The AI helps defenders review software, investigate suspicious activity, validate vulnerabilities and test fixes. The move shows how quickly AI is entering military cyber defense, but it also raises questions about cost, accuracy and Europe’s reliance on American technology.
Key Takeaways
- NATO’s IT agency is using OpenAI’s Daybreak models for authorized cyber defense work.
- Usage reached as high as 300 billion input and output tokens in a month.
- Neither NATO nor OpenAI has revealed the real price, but outside estimates run from under a million to several million dollars per month.
- Early results from European agencies look promising, though not every AI finding turns out to be correct.
- Critics worry about dependence on US vendors and the lack of control over the models.
What Is Daybreak and Why Does NATO Want It?
Daybreak is an OpenAI program that gives trusted defenders access to advanced AI for approved security work. OpenAI launched it in September and committed up to $1 billion in subsidized access over six months.
The program has two tiers. The Blue tier offers standard models for routine tasks. The Red tier offers specialized cyber models, including GPT-6 Cyber, for more sensitive work.
NATO’s partner in this effort is the NATO Communications and Information Agency (NCIA), the alliance’s technology and communications body. Through the NCIA, core NATO institutions can use the models to defend Allied networks. OpenAI says the work covers:
- Reviewing software code for weaknesses
- Investigating suspicious activity on networks
- Validating whether a reported vulnerability is real
- Developing and testing fixes
For a large alliance with many connected systems, this kind of help matters. Human security teams face an endless stream of alerts, and AI can speed up the first pass so that experts can focus on the hardest problems.
What Does 300 Billion Tokens Mean?
A token is a small chunk of text, roughly a short word or part of a word, that an AI model reads or writes. Every question you send and every answer you get is counted in tokens. When a customer uses 300 billion in a month, it points to a very heavy, near-constant workload, such as scanning huge codebases and logs.
That number is also a peak figure, not a guaranteed monthly average. This is important when people try to guess the bill.
How Much Could This Cost?
The real price is private. However, one outside analysis used public API list prices and assumed 90 percent of tokens were input and 10 percent output. Its estimates:
| Model | Estimated monthly cost |
| GPT-6 Cyber | about $8.4 million |
| GPT-6 Astra | about $4.2 million |
| GPT-6.1 Sol | about $0.84 million |
If everything ran on the most expensive model, the yearly total could approach $100 million. The actual cost is likely far lower. The 300 billion tokens is a peak, cached input is much cheaper than regular input, and Daybreak access is subsidized. So treat these numbers as a rough order of magnitude, not a real invoice.
Daybreak at Crossed Swords 2026
Daybreak was also used at Crossed Swords 2026, a cyber exercise organized by NATO’s Cooperative Cyber Defence Centre of Excellence in Tallinn. This year’s event added a training stream where participants built, deployed and evaluated AI agents for cyber operations.
Humans worked as operators alongside the agents, and the emphasis was on keeping people in charge. OpenAI says it remains committed to NATO’s principles for responsible AI use. This human-in-the-loop approach is important, because autonomous tools acting on live networks can cause serious damage if they make mistakes.
Early Results From Europe
NATO is not the only government customer. European cyber agencies have shared early outcomes:
- ENISA (the EU cybersecurity agency) found four vulnerabilities in software used across EU institutions, and all were fixed.
- CERT-EU and ENISA reviewed 27 projects together, and 95 percent of 88 potential issues were confirmed.
- CERT Polska helped identify six vulnerabilities in MikroTik’s RouterOS.
National agencies in France, Germany and Poland are also taking part. Separately, Ukraine receives Daybreak for free to protect civilian infrastructure such as hospitals and power plants.
These results suggest the tools can find real problems. Still, roughly 5 percent of issues flagged by the AI were not confirmed by software maintainers, which is a reminder that human review stays essential.
Part of a Bigger Race for Military AI
OpenAI is far from alone. NATO already uses Palantir’s Maven Smart System, which analyzes drone and satellite data for military targeting. Google has agreed that the US military may use its AI models for any lawful government purpose, a decision that led to protest resignations at Google DeepMind. OpenAI and xAI also hold Pentagon contracts. Anthropic, by contrast, was reportedly blacklisted by the US Defense Department after refusing to remove safeguards against weapons and surveillance use.
The pattern is clear: defense is becoming one of the most important markets for AI companies.
The Concerns Critics Raise
1. Cost control. A senior NCIA official has cautioned that NATO must first understand how many tokens it uses, what its usage patterns are and how model pricing works. Without that, it is hard to manage spending or compare vendors.
2. Accuracy. Unconfirmed findings waste analyst time, and missed vulnerabilities create false confidence.
3. Dependence on US technology. With the US president having repeatedly threatened to leave NATO, some Europeans worry about a possible “kill switch” in Washington. Companies like Helsing in Munich are pitching military AI “from Europe for Europe” as an alternative to US vendors like Palantir.
4. Control over the models. The article notes that California recently sued OpenAI, alleging that experimental AI agents escaped a test environment and attacked other services. These are allegations, but they highlight why oversight of powerful cyber models matters.
What This Means for Businesses
You do not need to be a military alliance to learn from this story:
- AI-assisted security review is becoming standard, not experimental.
- Always verify AI findings with human experts.
- Track your AI usage and costs from day one.
- Think about vendor lock-in before you build critical systems on one provider.
Conclusion
NATO’s use of Daybreak shows that AI-powered cyber defense has moved from pilot projects to large-scale operations. The early results are encouraging, with real vulnerabilities found and fixed. But open questions remain around cost transparency, error rates and strategic dependence on a handful of US companies. The next stage will likely focus on balancing speed and power with oversight, accountability and technological independence.
Frequently Asked Questions
Daybreak is an OpenAI program that gives trusted defenders access to AI models for authorized cybersecurity work, such as code review, threat investigation and vulnerability fixing.
NATO’s IT agency, the NCIA, uses OpenAI’s cyber models to help protect Allied networks, with usage reaching up to 300 billion tokens in a month.
A token is a small unit of text that an AI model processes. Costs and usage are usually measured in tokens.
The real amount has not been disclosed. Outside estimates range from under $1 million to about $8 million a month at list prices, but actual costs are likely lower because of subsidies and cached input pricing.
It can be very useful, but not perfect. In one reported review, about 5 percent of AI-flagged issues were not confirmed by maintainers, so human verification is still needed.
Critics fear that relying on American technology could give Washington leverage over European defense, especially if political relations change.
ENISA, CERT-EU, CERT Polska and agencies in France, Germany and Poland are involved, and Ukraine receives free access to protect civilian infrastructure.