What cheaper cache reads change in VulcanBench
A different cache-read rate cuts the estimated Sonnet sweep by 27.44%. The cost-per-success comparison tells a subtler story.
6 min readThe notebook
Notes with something
behind them.
Calculations, arguments, and the evidence underneath. Each note has a permanent home here.
↗
The denominator does the damage
Cheap per attempt can become ordinary per success. A closer look at the Medium-effort comparison.
3 min readA lower price is not a smarter model
What changes when you reprice a fixed workload—and which conclusions the arithmetic cannot support.
3 min readThe two explainers develop questions raised by the October 9 audit. They use the same frozen data; they are not additional benchmark runs.