What's this about?
swyx and Latent Space had early access and burned over 20 billion tokens of Astra on real AI-engineering work. Their verdict: it can choose and train models, label data, keep pipelines saturated, read logs, deploy and debug whole systems, and command subagents — at under $6 an hour.
What exactly is 20 billion tokens?
Roughly 15 million pages of text of work product — this is a review based on weeks of production-scale use, not an afternoon of prompting. The $6/hour figure divides total API cost by wall-clock hours of useful agent labor.
Why is this worth noting?
It's the most substantive early hands-on published, from reviewers with a track record — and "under $6/hour for competent ML-engineering labor" is the kind of number that reprices things. Calibrate for enthusiasm: early access selects for fans, as today's caveats note.