NVIDIA launches sol-pi: ai-driven efficiency boost for coding agents

NVIDIA, in collaboration with researchers from NTU and MIT, has introduced SoL-Pi, a set of efficiency mechanisms designed to enhance the performance of the open-source Pi coding agent. SoL-Pi reduces token traffic by up to 49% and cuts API costs by approximately 33% on the 51-task EdgeBench evaluation. The system, available on GitHub under the NVlabs license, is compatible with Pi 0.85.1 and Node.js 22.19 or newer. The innovation leverages auto-research loops to optimize the harness layer, which manages tool calls, context, and observations. Unlike traditional methods that focus on reducing cost per token, SoL-Pi minimizes the number of tokens consumed per task. The system maintains high performance, achieving 94.3% of Pi’s score on Opus 5 and 93.7% on GPT-5.6 Sol, while significantly lowering costs. The research team notes that the cross-model transfer is preliminary, with mechanisms triggering less frequently on Opus 5 due to the search process relying solely on GPT-5.6 Sol trajectories.