NVIDIA SoL-Pi cuts coding-agent tokens by up to 49%

NVIDIA’s SoL-Pi uses four harness changes to cut Pi coding-agent tokens by 44.7% to 49% on EdgeBench, though sources differ on quality.

NVIDIA researchers described SoL-Pi, a set of four harness-efficiency mechanisms for the Pi coding agent that MarkTechPost said was released as an open-source extension. The system reads agent traces and keeps changes that lower token use while aiming to preserve capability. MarkTechPost reported token-traffic reductions of 44.7% to 49.0% versus Pi, but its own copy conflicts on cost: one passage pairs that same range with API cost, while another says API cost fell roughly 33%. The Decoder is more specific on EdgeBench’s 51 public tasks, where the most efficient combination used 49% fewer tokens at 93.7% of Pi’s score, which is weaker than MarkTechPost’s claim of similar or better scores. Search spanned 535 environments and 152 directions with EdgeBench held out; the mechanisms fuse actions, compact context, pack observations, and shrink large logs. Results were mixed elsewhere: SoL-Pi solved 15 of 63 Terminal-Bench 4 CPU tasks versus 18 for Codex and Pi, GPT-5.6 Sol-trained mechanisms transferred to Opus 5 at 94.3% of Pi’s performance, and The Decoder estimated $8.75 to $13.50 saved per hour versus native Codex and Claude Code.

• NVIDIA researchers described SoL-Pi as four efficiency mechanisms for the open-source Pi coding agent; MarkTechPost called it an open-source extension.
• MarkTechPost reported token-traffic cuts of 44.7% to 49.0% versus Pi, while also stating API cost fell roughly 33% in one line and 44.7% to 49.0% in another.
• The Decoder said the most efficient four-mechanism mix used 49% fewer tokens on EdgeBench’s 51 public tasks and scored 93.7% of the original Pi harness.
• The four mechanisms fuse actions, compact context, pack observations, and reduce large logs.
• Search covered 535 environments and 152 directions and was held out from final EdgeBench evaluation.
• On Terminal-Bench 4, SoL-Pi solved 15 of 63 CPU tasks versus 18 for Codex and Pi.
• Mechanisms trained on GPT-5.6 Sol transferred unchanged to Opus 5 at 94.3% of Pi’s performance.
• The Decoder estimated savings of $8.75 to $13.50 per hour versus native Codex and Claude Code.