SoL-Pi cuts agent token traffic up to 49%

An NVIDIA-led team reports that open-source SoL-Pi cuts recorded token traffic 44.7%–49.0% and API cost by about one-third versus Pi across 51 EdgeBench tasks, according to the project page and the arXiv paper. The full stack retains roughly 94% of Pi’s average score, so the result is an efficiency trade, not free capability. Builders running long-lived coding agents should measure harness overhead before buying a larger model; Temporal’s agent-history cost analysis shows persistence can become its own bill once inference gets cheaper.