Nvidia harness cuts Pi coding agent token traffic by up to 49%

single source· 1 articles · confidence: medium · first seen 2026-09-22 05:04 UTC

What this means for you

The saving sits in the harness, not the model, which is the part you can change without waiting on a vendor. But nothing here is packaged for installation, no evaluation date is given, and the numbers are Nvidia's own — reproduce them before you plan around them.

Nvidia researchers have released SoL-Pi, four changes to the harness — the scaffolding around a model that manages prompts, tools and context — of the open-source Pi coding agent. The mechanisms were found by an AI running automated research loops across 535 environments. On EdgeBench, Nvidia reports token traffic down 44.7% to 49.0% and API cost down roughly 33%, while keeping about 94% of Pi's score on GPT-5.6 Sol and Opus 5. No evaluation date, benchmark version or harness link is given, and the work is reported by a single outlet from Nvidia's own figures.

Models in this story

Key facts

  • ·Nvidia researchers released SoL-Pi, four harness mechanisms for the open-source Pi coding agent. source
  • ·The mechanisms were discovered by an AI running auto-research loops across 535 environments. source
  • ·On EdgeBench, SoL-Pi cuts token traffic by 44.7% to 49.0%. source
  • ·API cost falls by roughly 33%. source
  • ·It retains about 94% of Pi's score on GPT-5.6 Sol and Opus 5. source

What the sources say

  • MarkTechPost — Reports the release, the search method that produced it, and the claimed savings.

Sources

The original reporting. Follow these — they did the work.

← the wire