NVIDIA's SoL-Pi uses AI agents to cut coding agent API costs by ~50% vs Codex
dr_cintas · x · 2026-10-03
NVIDIA's paper introduces SoL-Pi: a research AI agent reads the traces of a coding agent, finds where tokens are wasted, and proposes fixes; fixes that keep performance while lowering cost are retained each round. Four made the final harness:
- Action Fusion: merges actions into fewer steps
- Online Context Compact: clears finished work from context
- ObservationPack: shortens tool output
- Evidence-Preserving Reducer: delegates long reads to a cheaper model
On 51 held-out EdgeBench tasks with GPT-5.6 Sol and Opus 5: scores comparable to Pi at nearly half the tokens, saving $8.75–$13.50 per hour. Cuts API cost 50% versus Codex and 54.3% versus Claude Code.
More from coding & agent
- marimo launches open-source self-hosted notebook platform marimohub with pluggable storage, compute and SSO — S_Conradi · 2026-10-03
- xAI engineer says the team dogfoods Grok Bot daily, using bot to build bot — RachelVT42 · 2026-10-03
- Freebots gives every AI bot in its persistent 3D city free voice chat — Daniel_Farinax · 2026-10-03
- Dad turns Opus 5.5 game idea into live Roblox game with ~11,000 parts — minchoi · 2026-10-03
- Prime Sandboxes to add fork, checkpoint and restore for faster RL training — willcb · 2026-10-03
- SkyRL v0.4 trains 1T-param Kimi K2.7 with RL on just 16 B300 GPUs — casper_hansen_ · 2026-10-03