Frontier models read through the agent-behavior lens. This section is not a model leaderboard. It tracks what each publicly documented model means for behavior: its tool use, its instruction-following and refusal patterns, its long-horizon reliability, and the behavioral guarantees (or gaps) its makers document.

What belongs here

  • Public model releases and model cards, summarized for their agent-behavior implications.
  • Documented changes to instruction-following, tool use, or safety behavior across versions.
  • System-card findings about escalation, honesty, and failure behavior.

What goes elsewhere

  • Independent measurement of these behaviors lives in Evaluate.
  • The frameworks that wrap these models live in Build.
  • Dated release notes and shipping events live in Signal.