NVIDIA's SoL-Pi Cuts Coding Agent Token Traffic by ~45% and API Cost by a Third

songhan_mit · x · 2026-09-18

NVIDIA released the full SoL-Pi report (on Hugging Face), tackling a simple question: before scaling to thousands of parallel agents, can each agent waste fewer tokens?
Their approach is RSI-inspired auto-research loops at the harness layer: scale rollouts across increasingly diverse environments and let selection pressure find reusable improvements that transfer beyond their development setting.
Four mechanisms survived selection and form SoL-Pi: action execution, context compaction, observation handling, and delegated reading.
On the 51-task EdgeBench eval across GPT-5.6 Sol and Opus 5, SoL-Pi matches native Pi's performance while cutting recorded token traffic by 44.7-49.0% and API cost by roughly a third — estimated savings of $8.75-13.50/hour vs native Codex and Claude Code harnesses.

More from coding & agent