On 26 September 2026, THE DECODER detailed a Nvidia research system called SoL-Pi: instead of compressing the model, it optimizes the harness—the control layer between model and environment used by systems such as Codex or Claude Code. The core finding is economic: fewer tokens per task with similar performance on the held-out EdgeBench benchmark.
SoL-Pi automates what a human engineer usually does with long traces: a research agent watches runs, proposes harness changes, and keeps only candidates that hold capability while cutting cost. The authors frame the approach as recursive self-improvement loops, with a strict split between search and final evaluation.
Across 535 executable environments it explored 152 directions—495 tasks from GitHub issue–PR pairs and 40 synthetic cases—producing more than 3,000 runs and over 60,000 agent–environment interactions. More search does not guarantee better generalization; that is why EdgeBench was walled off from the search loop.
Of EdgeBench’s 51 public tasks, 11 once-validated finished candidates; 40 were reserved for final evaluation without feeding search. That methodological wall matters as much as the savings percentage: it fights the overfitting that has plagued auto-optimized harnesses in prior work.
Four mechanisms survived. Action Fusion merges two consecutive steps (e.g., edit code then run tests) into one model call. Online Context Compact trims accumulated context after each planning step when it can without losing critical information. ObservationPack archives long tool outputs and reinjects a short summary. Evidence-Preserving Reducer sends bulky logs to a cheaper model that extracts key findings with automatic verification.
On EdgeBench, the variant combining all four mechanisms uses 49% fewer tokens and reaches 93.7% of Pi’s score. Users who prioritize performance and pick only the strongest single mechanism beat Pi by 5.3% while still saving tokens. The aggregate range cited is 44.7–49% fewer tokens.
In dollars, the authors estimate roughly $8.75–$13.50 per hour saved versus native Codex and Claude Code harnesses, and $4.36–$5.71/h versus Pi, at then-current API prices. One efficient configuration cut an example cost from $1,339 to $894 with a slightly lower score.
The system was built with GPT-5.6 Sol and then applied unchanged to Opus 5, retaining 94.3% of Pi’s performance with similar savings, though mechanisms fired less aggressively—a sign the harness was biased toward Sol trajectories.
Outside EdgeBench the picture is messier. On 63 Terminal-Bench 4 CPU tasks, SoL-Pi solved 15 versus 18 for Codex and Pi, still ~25% cheaper than Pi. On Lean 4 tasks from IMO 2026, it solved three of six at the lowest cost per solved problem. A swarm of 20 SoL-Pi workers in a kernel-optimization test cut cost 26.8% versus a comparable Pi swarm.
Trade-offs exist: shorter contexts can reduce prompt-cache reuse. Context compression, warn studies cited in coverage, may preserve only a fraction of user instructions. More parallel sub-agents are not free either—they can burn tokens checking each other.
Neural Pulse reads SoL-Pi as evidence that the agent cost bottleneck is no longer only the frontier model’s price per token. The harness—how observations are packed, actions fused, and evidence summarized—can move the bill nearly as much as switching models.
For product teams, the implication is measurable: instrument traces, separate search benchmarks from evaluation, and test transfer across models before celebrating a lab “50% fewer tokens” claim on a single harness.
The paper does not claim a mathematical miracle or a Clay prize; it claims control-layer efficiency under held-out evaluation. That methodological honesty deserves copying in commercial releases that often blur score and cost.
Close: while OpenAI pauses tool-using workloads for safety, Nvidia shows another face of the same agent era—the war on wasted tokens. Whoever wins on harness can run more agents on the same budget, or the same work on a smaller bill.
