Another open model, you'll say. This one comes with one claim that, if it holds up, changes the math for anyone running agents in a loop, day after day.
Fact
The facts: on 8 July 2026, NVIDIA reported that Nemotron 3 Ultra - its open foundation model - achieves the highest accuracy among open models and runs at about 10 times lower inference cost compared to leading closed models. LangChain has optimized its agent framework Deep Agents specifically for Nemotron 3 Ultra, for long-running agentic workloads. Primary source: blogs.nvidia.com.
An agent doesn't ask the model once. It asks hundreds of times. In a loop, over and over. Under that kind of load, the cost of a single run stops being a small line item and becomes the ceiling - it decides whether you even get to production. Accuracy at an affordable price beats a peak you can't afford to run a thousand times.
The numbers come from NVIDIA. I take them with respect, not on trust - we measure them against our own tasks. If they hold up, an open model at that price changes what's reasonable to automate. The checking stays ours. The math gets lighter.
Official primary sources
NVIDIA - Nemotron Achieves Benchmark-Leading Performance With LangChain Deep AgentsThe visual is generated code-art, no third-party images.