NVIDIA Launches SoL-Pi to Cut Coding Agent Tokens by 49%
NVIDIA releases SoL-Pi on GitHub, cutting coding agent token traffic by up to 49% and API costs by over 33% on Pi 0.85.1.
Stock photo for illustration only, not from the actual event
- NVIDIA released SoL-Pi on GitHub under an MIT license.
- Cuts token consumption by up to 49% and lowers API costs by 33.5%.
- Runs on unmodified Pi 0.85.1 releases and Node.js 22.19 or newer.
- Employs fixed acceptance rules and a held-out EdgeBench evaluation suite.
Researchers at NVIDIA have released a new tool named SoL-Pi on GitHub as an MIT-licensed extension designed to run on unmodified Pi releases. The software has been tested and verified to work seamlessly with Pi 0.85.1 and Node.js version 22.19 or newer. While conventional efficiency optimization typically focuses on reducing the cost per token through faster kernels, quantization, or cheaper models, SoL-Pi takes a fundamentally different approach by reducing the actual number of tokens a task consumes through harness optimization.
Manually tuning a harness is a notoriously slow process where components are tightly coupled, meaning a fix in one area can inadvertently inflate costs in subsequent steps. Although systems like Meta-Harness attempt to automate this workflow, a recent study discovered that evolved harnesses frequently overfit their search tasks, yielding only marginal improvements on previously unseen tasks.
Coding agents often incur massive token traffic overhead during complex software engineering tasks, making harness management a critical bottleneck for development teams. By targeting token volume directly through automated research loops, NVIDIA's approach provides a viable pathway to substantially lower operational API expenses for enterprise-scale AI deployments.
The underlying system architecture relies on a research AI that observes execution traces generated by a separate agent running a base Pi instance. It then proposes targeted harness modifications and subjects them to rigorous testing. Each search operates as a disposable, isolated loop following an autoresearch cycle, extended with a Ralph Loop implementation step and an independent reviewer.
Stock photo for illustration only, not from the actual event
Acceptance rules are permanently frozen before any search commences, preventing the optimizer from altering them mid-process. Every capability metric must remain strictly within a predeclared tolerance threshold, and any candidate must demonstrate an improvement in at least one efficiency metric. Crucially, the EdgeBench suite remains held out from the search phase: out of 51 public tasks, 11 are allocated for one-way acceptance of frozen candidates, while the remaining 40 are reserved for final evaluation. Held-out results are never fed back into the search loop.
The full stack was initially constructed using GPT-5.6 Sol and subsequently migrated to Opus 5 without conducting additional searches. Operating on Opus 5, it retains 94.3% of the base Pi score while slashing token traffic by 44.7% and API costs by 33.5%. When evaluated on GPT-5.6 Sol, it preserves 93.7% of the original score while achieving 49.0% fewer tokens and a 33.2% reduction in overall cost. Furthermore, performance-focused configurations deploy the optimal single mechanism for each respective backend—ObservationPack on GPT-5.6 Sol and Action Fusion on Opus 5—boosting scores by 5.3% and 12.8% above the baseline Pi.
Although the full stack on GPT-5.6 Sol experiences an increase in cache-write traffic from 0.0141 billion to 0.0316 billion tokens, total financial cost still drops significantly from $1,339 to $894. The research paper estimates hourly savings ranging from $8.75 to $13.50 compared to native Codex and Claude Code harnesses, and $4.36 to $5.71 compared to standard Pi. Nevertheless, the research team notes that cross-model transfer remains preliminary, as the mechanisms trigger less frequently on Opus 5, potentially because the underlying search relied exclusively on GPT-5.6 Sol trajectories.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment