NVIDIA Open-Sources SoL-Pi Efficiency Layer, Cutting Token Usage by Up to 64%

On September 11, NVIDIA announced the open-source release of SoL-Pi, an efficiency enhancement layer built on the Pi model. It aims to reduce token consumption and API call costs in AI agents while enabling AI to iteratively optimize itself.

SoL-Pi incorporates four core mechanisms: action fusion combines an edit and its subsequent commands into a single local execution sequence, skipping an intermediate model decision step; online context compaction dynamically evaluates when to compress context by subtask, ensuring rewrite costs are covered by future savings; output archiving with indexing stores full tool outputs on local disk, keeping only a short handle and excerpt in context for on-demand retrieval; and an evidence-preserving reducer delegates the first pass through long logs to a low-cost model, with outputs verified against archived logs before being passed to the main model.

According to official benchmarks, SoL-Pi uses 45%-49% fewer tokens than the base Pi and reduces costs by about one-third. Compared with the default harnesses of Codex and Claude Code, it cuts token usage by 35%-64% and API call costs by 50%-54%, saving $8.75-$13.50 per hour in specialized research scenarios.

The project was led by NVIDIA's Efficient AI team under research director Song Han, and the code is available on GitHub (NVlabs/SoL-Pi), installable into existing Pi setups with a single command.