Vera 1.6 pushes agentic context to 1M

The interesting part here is not just the million-token headline but the fact that Cortex is optimizing the stack around autonomous tool use and long-horizon execution, which is where many general-purpose models still wobble.

• –Cortex is making a compute-efficiency bet with a 3:1 mix of linear and quadratic attention plus 256 experts with only 9 active per token, aiming to make million-token agent loops deployable instead of just demoable.
• –The benchmark spread is strong in math, multilingual tasks, document understanding, and video, but the 41.6% Terminal-Bench 2 result shows terminal-heavy agents are still the weakest link.
• –A 150B-token synthetic training corpus, agent-specific RL, and a broader platform push around integrations and agents suggest Cortex is targeting enterprise agent systems rather than pure model leaderboard hype.