NVIDIA Nemotron 3 Ultra tops open models on LangChain Deep Agents at 10x lower cost

Nemotron 3 Ultra is NVIDIA's argument that open-weight models can now match closed frontier models on real agentic work while radically undercutting them on cost. On LangChain's Deep Agents harness—a framework for multi-step tool-using agents—the model tops open competitors and reaches parity with leading closed models on business tasks at roughly one-tenth the inference cost per run, completing more tasks at higher throughput.

Mechanically, LangChain built a tuned harness profile specifically for Nemotron 3 Ultra to extract that performance, which matters because agentic accuracy depends heavily on how a model is scaffolded (tool interfaces, feedback loops), not just raw weights. The 10x cost claim is measured per run on comparable business-task parity, not a raw token-price comparison.

Competitively, this feeds the week's dominant cost theme—Microsoft routing to MAI, Chinese open models undercutting on price, DeepSeek's chip play—and positions NVIDIA (which sells the GPUs everyone runs on) as also a purveyor of cost-efficient open models, hedging both sides of the market. It complements AWS's Swami Sivasubramanian research on model-agnostic agent harnesses (Simple Strands Agent) posted the same week. Skeptics will note vendor-run benchmarks flatter the vendor and that a tuned harness isn't representative of default deployments. Watch third-party reproductions and whether enterprises actually swap closed models for Nemotron on agent workloads.