The cost of each study varies with its scale and complexity, but my first month of running JackBench came to 760 USD. Below are examples of what some of the individual learnings cost.

Table 1 · Provider spend by study
ItemAmount (USD)Notes
Learning that GPT-5.6 Sol gains 19.4 points in the right harness 153 7,255 provider requests
Learning how Claude Code, Codex and Hermes handle the same tasks ~170 finished under a 500 USD ceiling
Learning what a managed agent product (Warp) does with those tasks 28 cost as reported by Warp
Learning that local Qwen 3.6 27B performs better in Pi 0 run on my own hardware, no provider bill

Hosted figures come from sealed per-attempt receipts; the Warp figure is as Warp reported it.

What a donation buys

I'll keep sharing my research for free. I'll test frontier models as they arrive, and I'll run more experiments locally on anything up to 230B parameters. I've got a backlog of studies I want to run for my own products, but at some point I may design and run studies at the request of donors. I do this for fun, but it is expensive. Anyone who donates, I truly thank you from the bottom of my heart, and I hope my work gives something back to you.