NVIDIA Labs' SoL-Pi harness cuts coding-agent costs by up to $13.50 an hour versus native Codex and Claude Code

A dozen agents run over a long weekend, and the API invoice on Monday has no column for the work the swarm did twice. SoL-Pi is an attempt to put a figure in that column, and the figure comes in two parts.

Divide the two ranges the page gives. $4.36 against $13.50 is 32 percent. $5.71 against $8.75 is 65 percent. So between roughly a third and two thirds of the hourly saving against the native harnesses comes from what the search loops added on top of Pi ^[19]. The remainder comes from being on Pi at all. The page says the range reflects the model backend and does not pair the endpoints, so those two percentages are the outer bounds ^[4]. Both comparisons are harness against harness, and capturing either saving means moving agents onto a different harness.

The reason for doing efficiency work before scaling recursive self-improvement, in the page's own telling, is budgetary. RSI "is itself token-intensive: every attempt to produce a better system spends tokens, whether or not it succeeds," the writeup says ^[9]. It leaves the ordering as an open question, asking whether AI should first be asked to make AI more efficient ^[10].

The part of the method that makes the efficiency claim checkable is the split. Candidates are developed and screened on a training set, then frozen along with an acceptance rule before a held-out test set decides them. Held-out trajectories never re-enter analysis, and no agent inside the auto-research loop sees held-out results ^[15]. EdgeBench, a 51-task suite of long-horizon executable agent work, is reserved for that final evaluation ^[6]. The search itself optimises cost or token use subject to a capability-preservation criterion declared in advance ^[11]. NVIDIA did not publish how many of the proposed directions survived the first pass, or what that criterion was set to ^[20].

A narrower test than the headline range: take one long run you have already paid for and work out what share of its tokens went to re-reading context the agent had already seen. If that share is small, the harness is not where your money is going. If it is large, the page gives you the unit to measure it in: one researcher, one problem, one hour, at official API-equivalent pricing ^[4]. A team on negotiated rates or its own hardware pays something else and has to redo the number.