What happens when the most capable AI models aren't broadly released but reserved for enterprises that can afford premium access?
Anthropic's Project Glasswing is making that a concrete question. The cybersecurity-focused model launched under tight controls: invite-only access, premium pricing, selected partners, and an emphasis on enterprise deployment rather than open availability.
That's a notable departure from how AI tools have mostly been released since 2022. The competitive dynamic has largely pushed companies toward capability-at-commodity-price - OpenAI, Anthropic, and Google cutting access costs to stay ahead of each other. Glasswing suggests a parallel track: top-tier specialized models sold to enterprise clients with access controls and pricing to match.
The Security Case for Restricted Access
Cybersecurity is one of the few domains where the argument for restricting AI access holds up under real scrutiny.
A model capable enough to help security teams find vulnerabilities, analyze malware, and build defensive tools is the same model that could help attackers do those things. The dual-use problem is built into the domain. Giving anyone with a credit card access to a world-class offensive security AI is a different risk calculation than giving them access to a writing or summarization tool.
Anthropic's controls on Glasswing - selected partners, verified use cases, enterprise contracts - create accountability layers that open access wouldn't. Whether that's enough to prevent misuse is a legitimate debate, but the logic isn't just commercially convenient. It's a reasonable response to a genuine problem.
Specialized AI Going Behind Paywalls
The counterargument is access concentration. If the most capable specialized models go to large enterprises first at premium prices, smaller security teams - at mid-market companies, nonprofits, government agencies with constrained budgets - are left working with less capable tools.
In security, that gap has real consequences. A smaller team using a weaker model might miss what a larger competitor's enterprise model would catch. The divide between well-resourced and under-resourced security operations could widen if AI capability becomes another thing money buys.
Glasswing may also preview how frontier AI companies monetize their most capable models going forward: commodity pricing for general-purpose AI, premium restricted access for highly capable specialized models in sensitive domains. Every enterprise software category eventually moves toward tiered pricing. The question for practitioners is whether Glasswing's invite-only access opens up over time and at what cost. The broader question is whether this pattern becomes the template for future releases in legal AI, medical AI, and financial AI - domains where the dual-use concern and liability exposure are similarly high.
The "AI for everyone" framing that drove the 2022-2024 wave is quietly giving way to something more stratified.