NVIDIA’s SoL-pi slashes AI token costs with recursive loops
The gist
NVIDIA and KuCoin’s open-source SoL-Pi harness just rewrote the rules on AI efficiency—slashing token use and API bills with a recursive, self-optimizing system.
What to know
- SoL-Pi launched in September 2026, boasting up to 64% fewer tokens and 50–54% lower API costs compared to previous Codex and Claude Code harnesses.
- Its recursive generate-test-prune loops let agents run hundreds of experiments autonomously—700 in two days—keeping only genuine improvements.
- The result: cost per task dropped from $0.21 to $0.12, median completion time fell by 44%, and harness optimization became a repeatable, automated process.
A New Era in Harness Design
SoL-Pi’s open-source debut marked a pivotal shift by automating harness optimization through recursive AI-driven searches, transforming efficiency from a one-off tweak into a repeatable, data-backed process.
The dated break came in September 2026, when NVIDIA moved SoL-Pi from concept to public release, open-sourcing the harness on GitHub and framing it explicitly as a harness-level efficiency upgrade aimed at cutting token and API costs. That release was reinforced days later by the September 2026 arXiv publication, “SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness” (https://arxiv.org/abs/2609.20519), which positioned the system as an RSI-inspired framework rather than a one-off prompt tweak or isolated engineering patch.
What made the release an inflection was not just availability but the headline economics attached to it: SoL-Pi’s reported performance showed “Token consumption reduced by up to 64%, and API call costs dropped by 50%-54%.” KuCoin’s explainer tied that upper bound directly to comparisons against earlier harnesses, stating, “Compared to the original harness for Codex and Claude Code, it uses 35% to 64% fewer tokens and has a marked cost reduction of 50% to 54%,” making the cost case legible to operators evaluating deployment budgets.
The evidence for those savings was presented as the product of a recursive auto-research process rather than a single manual redesign, which is why SoL-Pi marked a dated turn in how harness efficiency was discussed. As the explainer described it, “The AI's entire search began with 152 candidate directions and ultimately narrowed down to four killer architecture mechanisms,” a concrete signal that NVIDIA and KuCoin were showcasing an automated search-and-refinement pipeline capable of producing measurable cost reductions at harness level.
Recursive Loops Drive Relentless Gains
By layering generate-test-prune cycles with a self-improving outer loop, SoL-Pi turned harness refinement into a continuous experiment engine—achieving hundreds of real improvements while slashing costs and run times.
Recursive loops become an efficiency engine when the harness itself formalizes the generate-test-prune cycle. An agent proposes a change, a verifier grades it, state records failures and survivors, and rollback discards bad edits; “that budget yields roughly 12 experiments per hour, so about 100 run overnight,” while “the loop halts when the goal is met, or after N attempts,” making harness refinement a repeatable operating process rather than a manual sprint.
The recursive step adds an outer loop that improves the search process itself: it watches where the inner loop stalls, writes new mechanisms, injects them, and reruns, which is why the outer loop cut val_bpb 5x more than the single loop (-0.045 vs -0.009). Guardrails matter because “that separation stops the agent from making the test easier instead of the model better,” and in practice the system “ran for two days and completed about 700 experiments, keeping 20 genuine improvements,” while harness optimization also reduced cost per task from $0.21 to $0.12, cut median completion time 44%, held quality near parity or improved, and made TrueForge about 2.7x cheaper for the same result with roughly 40% fewer trips back to the model.
