Nvidia released the SoL-Pi system on September 26, 2026, saying it optimizes the harness of coding agents to cut token usage nearly in half. It is the company's first major update to its coding agent framework since its initial release in 2025.
Nvidia reported a 50 percent reduction in token usage compared to Codex and a 54.3 percent reduction compared to Claude Code on EdgeBench, measured on the EdgeBench benchmark. That compares with the original Pi harness, which had higher token costs.
The SoL-Pi system is built on the Pi harness and targets coding agents used in environments like Codex, Claude Code, or OpenClaw. Availability begins with public access, initially for developers and researchers.
"SoL-Pi saves 50 percent compared to Codex and 54.3 percent compared to Claude Code on EdgeBench," said Jonathan Kemper, the paper's author. The system automatically optimizes the harness by analyzing agent traces and proposing changes that maintain performance while cutting costs.
The announcement follows a growing trend in AI research focused on reducing the computational and financial costs of large language models. Nvidia's approach highlights the importance of optimizing the control layer rather than just the model itself.
Nvidia did not say how the system performs on benchmarks beyond EdgeBench, and the researchers raised concerns about overfitting to training tasks. The company also noted that the system's efficiency gains come with trade-offs, such as reduced context retention.
Source: thedecoder
