Each AutoResearch experiment has costs: the LLM API call (your agent reasoning about what to try) and the GPU time (training for minutes).
The API cost is small. The agent reads train.py, results.tsv, and program.md, reasons about a change, and outputs edited code. That's a few thousand tokens per cycle. At Claude Opus pricing, each cycle costs a fraction of a cent.
The GPU cost dominates. An H runs about \25$0.174090$0$. You already own the hardware.