NVIDIA's SoL-Pi Nearly Halves Agent Token Traffic

The important result is not another smarter model; it is a better control loop around existing ones. Harness engineering is becoming a serious optimization layer, though the benchmark scope leaves room for skepticism.

• –NVIDIA searched roughly 150 proposed mechanisms across 500 environments, 3,000 runs, and 60,000 agent-environment interactions.
• –The four retained techniques target predictable waste: repeated validation turns, oversized observations, long diagnostic logs, and stale context.
• –SoL-Pi transferred from GPT-5.6 Sol to held-out Opus 5 without further adaptation, retaining comparable performance while reducing traffic and cost.
• –Results are not universally dominant: on Terminal-Bench 4, SoL-Pi solved 15 tasks versus 18 for both Codex and Pi.
• –The MIT-licensed extension uses Pi's public APIs and remains opt-in, but its remote log-reduction feature deserves careful security review.