AI Foundations · The 2027 model lineup
You can describe what changed when Anthropic announced Mythos Preview and Project Glasswing, and explain why "staged release" is now a model-tier category rather than a marketing phrase.
Refreshed: May 2026.
Through 2024 and 2025, every major lab followed the same product cadence: train a flagship, run safety evaluations, ship it through a consumer chat interface and a developer API. Capability and access moved together. A more capable model meant a model more people could use.
That pattern broke in spring 2026. Anthropic's own Mythos Preview page said the lab was testing a new model class that it did not plan to release as a normal public product. The model is called Mythos (the generation, analogous to "Claude 4"), and its first tier is called Capybara. Anthropic's framing: capability "currently far ahead of any other AI model in cyber capabilities," with autonomous discovery of novel zero-days, including a 17-year-old FreeBSD remote-code-execution flaw.
The access mechanism was named Project Glasswing — a partnership with roughly 50 critical-infrastructure organizations (AWS, Apple, Cisco, Google, JPMorgan, Microsoft, Nvidia, CrowdStrike, Palo Alto Networks, the Linux Foundation, ~40 more) plus $100M in usage credits and $4M in donations to open-source security. No public API. No consumer chat. No fine-tuning. No download.
Chapter contains 3 lessons.