Frontier AI access is shrinking, not expanding, as security and compute pressures mount

The popular assumption that frontier AI tokens will become cheap and universally available is breaking down. Anthropic’s recent Mythos rollout — a top-tier cybersecurity model offered only to a handful of vetted U.S. partners — and OpenAI’s similarly gated Daybreak initiative signal that leading labs are abandoning broad release in favor of curated distribution. The U.S. government appears poised to formalize this pattern, viewing selective access as both a misuse-mitigation tool and a way to preserve offensive options around newly discovered vulnerabilities.

Three compounding forces drive the squeeze. Security concerns push labs to gate dangerous capabilities, particularly in cyber and bio domains, while fears of model theft and API-based distillation by fast followers like DeepSeek are prompting tighter KYC and geopolitically conditioned access. Compute is the second crunch: unlike software, frontier inference carries high marginal costs, and efficiency gains apply to last year’s capabilities, not the current frontier — so each new tier gets more expensive, not less. Anthropic is reportedly negotiating capacity deals with rivals including xAI to keep up.

The implication is uncomfortable for actors outside the U.S. inner circle: middle powers, allied governments, and most enterprises betting on “good enough” models will find themselves structurally behind. Defenders will not reliably get parity with attackers, and competitive strategies premised on universal access to the best AI need rethinking.