What's this about?
OrangeAI's substantive read: Astra's intelligence profile is unusual — far fewer reasoning steps for more intelligence per token, big science gains, small coding gains, still behind Fable on code.
What exactly is "intelligence per token"?
How much useful cognition each generated token carries. A model needing fewer written steps for the same conclusion thinks *denser* — consistent with the recurrent-depth architecture doing work silently (with the monitorability costs covered above).
Why is this worth noting?
It matches the external evidence (Coding Agent Index, Cognition's numbers) and sharpens the market: Astra for science and computer-use, Fable for code — capability profiles diverging by domain rather than one model winning outright.