Anthropic's Mythos Vault Strategy Proves Frontier Model Lockdown Is Here
TexTak places frontier open-source model parity at 69%, but Anthropic's decision to vault Claude Mythos Preview—its most capable model yet—represents the clearest signal that leading AI companies are breaking from the open paradigm. While Meta continues pouring resources into open-source and Chinese models dominate benchmarks, the cybersecurity capabilities that forced Mythos into restricted access suggest frontier labs are developing capabilities they fundamentally cannot release publicly.
The Mythos vault decision is a structural shift, not a one-off safety theater. Anthropic built Project Glasswing—a coalition of 40 organizations including AWS, Apple, Microsoft, and Google—specifically to contain a model too dangerous for public release. This isn't the gradual capability withholding we've seen before; it's the first time a frontier lab has acknowledged that their latest model crosses a line that makes open release impossible. The fact that it's cybersecurity capabilities—finding thousands of zero-day vulnerabilities across major operating systems—that triggered the vault tells us frontier labs are developing dual-use capabilities that dwarf anything in the open-source ecosystem.
Our 69% forecast for open-source parity assumes that capability gaps can be bridged through compute, data, and technique sharing. But if frontier models are developing capabilities that are structurally unreleasable—whether cybersecurity exploits, bioweapon design, or other dual-use applications—then the apparent benchmark convergence between open and closed models becomes meaningless. Chinese models like GLM-5.1 and Qwen3.5 may dominate public benchmarks, but they're competing against deliberately hobbled versions of what frontier labs actually possess. When American companies adopt Alibaba's Qwen for chatbots, they're not accessing frontier capabilities—they're accessing the frontier labs' public-facing performance ceiling.
The counterargument centers on Meta's continued commitment to open-source AI, with massive investment in the Llama ecosystem and stated philosophical opposition to closed development. Meta's approach could force other labs to compete in the open, making vault strategies unsustainable. Additionally, the specific cybersecurity capabilities that triggered Mythos's vault may not represent general intelligence advancement—it could be a narrow, specialized capability that doesn't affect the broader race for model parity.
What we're potentially underweighting is the possibility that frontier labs have already developed capabilities far beyond what they're showing publicly, making our parity forecast not just wrong but obsolete. The Mythos vault could be the tip of the iceberg—if leading labs are sitting on models with step-change capabilities they can't release, the open-source community isn't closing a gap but chasing a mirage. We'd drop below 50% if two more frontier labs vault their next models over safety concerns, or if leaked benchmarks show vaulted models performing 2+ standard deviations above the best open models.