Five separate outlets ran essentially the same headline this week: OpenAI and Anthropic are in a price war, and cheaper Chinese rivals are the reason. The Financial Times, TechRepublic, Moneycontrol, YourStory, and Business News Nigeria all landed on the same story within days of each other. When coverage converges that fast, something structural is shifting, and I want to talk about what that shift means for the people who defend AI systems rather than the people who sell them.
I’m Omar Hassan, and I spend my days looking at how AI bots get attacked. So when I read that frontier labs are slashing prices because open-weight models from China are gaining ground, I don’t think about margins or market share first. I think about attack economics.
Every attack has a budget
Security has always been a cost equation. Attackers do things when the expected payoff exceeds the expected cost, and defenders survive when they make attacks expensive enough to not be worth running. For the past two years, one of the quiet brakes on AI-powered abuse has been price. Running large-scale prompt injection campaigns, automated social engineering, or bot-driven scraping through a frontier API cost real money. Not prohibitive money, but enough to keep the casual abuse tier thin.
A price war removes that brake. When OpenAI and Anthropic cut costs, they cut costs for everyone, including the people probing your chatbot for jailbreaks at scale. Cheaper inference means an attacker can run more variations, more attempts, more automated reconnaissance against your AI-facing surfaces for the same spend. The defender’s job gets harder in direct proportion to how affordable the attacker’s tooling becomes.
Open weights change the defense calculus too
TechRepublic’s framing is the one I keep returning to: the pressure isn’t just from cheaper hosted rivals, it’s from open-weight models gaining ground. That matters for security in two directions at once.
On the attacker side, open weights mean no API, no rate limits, no usage policies, no trust and safety team reviewing your account. An adversary running a capable model on their own hardware answers to nobody. The abuse mitigations that hosted providers build into their platforms simply don’t exist in that world, and every improvement in open-weight capability widens what a fully unmonitored attacker can do.
On the defender side, though, there’s a real upside that doesn’t get enough attention. Cheaper models, whether hosted or open-weight, make defensive AI affordable. Classifying inbound prompts for injection attempts, screening outputs before they reach users, running adversarial testing against your own bots before deployment — all of this is inference-heavy work that many teams have skipped because the cost didn’t pencil out. A price war changes that math in our favor too.
What a race to the bottom does to safety spending
My bigger worry is organizational. Price wars compress margins, and compressed margins force choices about where money goes. Safety research, red-teaming, abuse monitoring, and incident response are cost centers in a spreadsheet, even though they’re the reason enterprise customers trust these platforms at all. I’m not claiming any lab has cut safety spending — I have no evidence of that. But the incentive pressure is now real and pointed in a direction that should make anyone in this field pay attention.
There’s also a procurement effect worth watching. When prices drop across the board, companies that previously couldn’t justify AI integration suddenly can. That means a wave of new AI deployments built by teams with no prior experience securing them. In my experience, the most vulnerable AI systems aren’t the ones built by the labs. They’re the ones built quickly, cheaply, and enthusiastically by teams who assumed the model provider handled security for them.
What defenders should do now
- Rebudget your defensive inference. If model costs are falling, revisit the defensive layers you priced out six months ago. Output screening and prompt classification may now be affordable.
- Assume attacker volume goes up. Cheaper tokens mean more automated probing. Rate limiting, anomaly detection, and abuse telemetry on your AI endpoints matter more than they did last quarter.
- Plan for open-weight adversaries. Defenses that rely on provider-side safety filters won’t hold against attackers running unrestricted local models. Build your own layer.
- Audit the new deployments. If falling prices trigger a rush of AI features inside your organization, get security review into that pipeline before launch, not after the first incident.
Cheap intelligence is coming to everyone, on both s
đź•’ Published: