Discussion about this post

User's avatar
Stas's avatar

Spot on analysis, Anton. The shift away from paying "System Two" prices for fast micro-decisions means we can finally let code own the control flow again.

JP's avatar

The metering angle is the useful part of this piece. Parallel typed questions with free output tokens really does change how you price agent loops, and you are right that calibrated 0.8 is a long-run rate, not a guarantee on any one call.

I'd push on the unreproduced frontier-parity claim a bit harder in practice: treat vendor laptop numbers as a starting point, then pin accept/reject thresholds on your own held-out traffic, including the middling band you still send to a System Two model. I run JevMade as a hub for guides and experiments on this stack if useful.

Ready for more?