{"version":1,"type":"story","url":"https://digestai.news/story/nvidia-releases-sol-pi-cutting-coding-agent-token-traffic-by-up-to-49","json":"https://digestai.news/story/nvidia-releases-sol-pi-cutting-coding-agent-token-traffic-by-up-to-49.json","markdown":"https://digestai.news/story/nvidia-releases-sol-pi-cutting-coding-agent-token-traffic-by-up-to-49.md","slug":"nvidia-releases-sol-pi-cutting-coding-agent-token-traffic-by-up-to-49","headline":"Nvidia releases SoL-Pi, cutting coding agent token traffic by up to 49%","summary":"Researchers from Nvidia, NTU and MIT have open‑sourced SoL‑Pi, an MIT‑licensed extension for the Pi coding agent. The four efficiency mechanisms—Action Fusion, Online Context Compact, ObservationPack and Evidence‑Preserving Reducer—were discovered by an auto‑research loop that explored 152 directions across 6 families in 535 environments, running more than 3,000 trials and 60,000+ agent‑environment interactions.\n\nOn the 51‑task EdgeBench suite, SoL‑Pi reduces token traffic by 44.7% to 49.0% compared with vanilla Pi and lowers API cost by roughly 33%, while retaining about 94% of Pi’s average score (94.3% on Opus 5, 93.7% on GPT‑5.6 Sol). Additional benchmarks show lower costs on Terminal‑Bench 4 and comparable problem‑solving on IMO 2026 Lean 4‑verified tasks. The code runs on an unmodified Pi 0.85.1 release with Node.js 22.19+ and is available on GitHub.","keyPoints":["SoL-Pi reduces EdgeBench token traffic by 44.7%‑49.0% versus Pi, lowering API cost about 33%.","The four mechanisms keep roughly 94% of Pi’s performance while running on GPT‑5.6 Sol and Opus 5.","SoL-Pi is an MIT‑licensed GitHub extension for Pi 0.85.1, tested with Node.js 22.19+, released by Nvidia, NTU and MIT researchers."],"whyItMatters":"Reducing token traffic and API cost makes long‑running coding agents affordable, enabling more complex software development automation and broader adoption of AI‑driven coding tools.","category":{"slug":"agents","name":"Agents & Tools","url":"https://digestai.news/category/agents"},"entities":{"companies":["Nvidia","NTU","MIT"],"models":["Pi","GPT-5.6 Sol","Opus 5","Codex","Claude Code"],"people":["Asif Razzaq"]},"firstPublishedAt":"2026-09-22T05:04:43Z","updatedAt":"2026-09-22T05:04:43Z","sourceCount":1,"hasPrimarySource":false,"sources":[{"outlet":"MarkTechPost","title":"NVIDIA Introduces SoL-Pi: Auto-Research Loops That Cut Coding Agent Token Traffic by Up to 49%","url":"https://marktechpost.com/2026/09/21/nvidia-researchers-have-released-sol-pi","publishedAt":"2026-09-22T05:04:43Z","type":"press","primary":false,"lead":true}],"sourceNotes":null,"discussions":[],"thread":null,"cite":{"text":"Digest AI, \"Nvidia releases SoL-Pi, cutting coding agent token traffic by up to 49%\", 22 September 2026, https://digestai.news/story/nvidia-releases-sol-pi-cutting-coding-agent-token-traffic-by-up-to-49","publisher":"Digest AI","title":"Nvidia releases SoL-Pi, cutting coding agent token traffic by up to 49%","datePublished":"2026-09-22T05:04:43Z","url":"https://digestai.news/story/nvidia-releases-sol-pi-cutting-coding-agent-token-traffic-by-up-to-49"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}