Nvidia's SoL-Pi system cuts coding agent token usage nearly in half by optimizing the harness
Nvidia's SoL-Pi research system cuts coding-agent token use nearly in half by optimizing the harness.
Nvidia researchers describe SoL-Pi, a system that reads coding-agent traces and keeps harness changes that maintain capability while lowering token cost. On EdgeBench's 51 public tasks, the most efficient combination of four mechanisms used 49% fewer tokens and scored 93.7% of the original Pi harness, with estimated savings of $8.75 to $13.50 per hour versus native Codex and Claude Code. The search covered 535 environments and 152 directions and was walled off from final EdgeBench evaluation. Results were mixed elsewhere: SoL-Pi solved 15 of 63 Terminal-Bench 4 CPU tasks versus 18 for Codex and Pi, and mechanisms trained on GPT-5.6 Sol transferred to Opus 5 at 94.3% of Pi's performance.
- SoL-Pi proposes harness changes and keeps only those that preserve performance while cutting tokens.
- The most efficient variant uses 49% fewer EdgeBench tokens at 93.7% of Pi's score.
- Four mechanisms fuse actions, compact context, pack observations, and reduce large logs.
- Search spanned 535 environments and 152 directions, with EdgeBench held out from feedback.
- Unchanged transfer to Opus 5 retained 94.3% of Pi performance with similar savings.
- Vendors
- Nvidia
- Products
- SoL-PiCodexClaude CodePiOpenClaw
- AI models
- GPT-5.6 SolOpus 5DeepSeek V4 Flash
From the original article
5 min read at the-decoder.com
SoL-Pi cuts coding agents' token usage by up to 49 percent with little change in performance by optimizing the control layer between the model and its environment. A research agent tested 152 approaches across more than 3,000 runs to develop the system, though the gains were smaller on other benchmarks.
By Jonathan Kemper · the-decoder.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.