NVIDIA researchers have launched SoL-Pi, an open-source extension available on GitHub under an MIT license. The tool is compatible with Pi release 0.85.1 and requires Node.js 22.19 or newer. Unlike methods that rely on quantization or faster kernels, SoL-Pi focuses on reducing the total number of tokens consumed by coding agents during task execution.
The system utilizes an automated research loop to optimize harness configurations, which manage tool calls and context. During testing on the GPT-5.6 Sol model, SoL-Pi reduced token traffic by 49.0% and lowered API costs by 33.2%, while maintaining 93.7% of the original performance score. On the Opus 5 model, the tool achieved a 44.7% reduction in token usage and a 33.5% decrease in costs, retaining 94.3% of the baseline score.
The evaluation process involved 51 public tasks from EdgeBench, with 11 tasks used for acceptance and 40 for final assessment. While the stack increased cache-write traffic on GPT-5.6 Sol, the total operational cost dropped from $1,339 to $894. Researchers estimate hourly savings ranging from $8.75 to $13.50 compared to native Codex and Claude Code harnesses.
The research team noted that cross-model transfer remains in a preliminary stage. Mechanisms such as ObservationPack and Action Fusion were identified as top-performing components for specific backends, lifting performance scores by 5.3% and 12.8% respectively over the base Pi implementation. The optimizer operates under fixed acceptance rules to ensure capability metrics remain within predefined tolerances.
