NVIDIA’s NVLabs has open-sourced SoL-Pi, an efficiency layer designed for the coding agent harness Pi. The project used an automated research workflow in which multiple agents reviewed Pi’s execution traces, proposed changes, rewrote code, and verified results automatically. Out of 152 initial optimization directions, the team kept four, all aimed at reducing repeated overhead. The changes include continuing straight into testing after code edits to avoid an extra model call, storing previously read large files and long outputs locally for later retrieval, deciding after subtask completion whether context compression is needed, and using a cheaper model to extract key points from very long logs before checking them against the original log. On EdgeBench, SoL-Pi used 45% to 49% fewer tokens than the base Pi and reduced cost by about one-third. Against the native harnesses of Codex and Claude Code, token usage fell by 35% to 64%, while estimated costs based on API list prices were 50% to 54% lower. The trade-off was a modest drop in capability, with SoL-Pi retaining about 94% of Pi’s average score.
NVIDIA NVLabs has open-sourced SoL-Pi, adding an efficiency layer to the coding agent harness Pi. The research team built an automated workflow in which multiple agents analyzed Pi’s execution traces, proposed optimizations, rewrote code, and verified the results automatically.
The process began with 152 candidate directions. In the end, the team kept only four, and all four were aimed at cutting repeated overhead.
Four retained changes targeted repeated overhead
According to the release, the four optimizations were:
- moving straight into testing after code changes, which removes one model call;
- storing previously read large files and long outputs locally and retrieving them only when needed;
- deciding whether to compress context after a subtask is finished;
- sending very long logs to a lower-cost model first to extract key points, then checking that the content does in fact come from the original logs.
EdgeBench results showed lower token use and lower cost
On EdgeBench, SoL-Pi used 45% to 49% fewer tokens than the original Pi, with costs about one-third lower. Compared with the native harnesses used by Codex and Claude Code, SoL-Pi reduced token usage by 35% to 64%. Based on API list pricing, costs were 50% to 54% lower.
Lower cost came with a small performance trade-off
The savings came with some loss of capability. SoL-Pi retained about 94% of Pi’s average score.
On the 63 CPU tasks in Terminal-Bench 4, SoL-Pi solved 15. Codex and Pi each solved 18.
This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan. Disclaimer:
The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.
Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.
