NVIDIA Nemotron open models deliver 20x cheaper inference than frontier; Harvey, Glean ship post-trained agents

NVIDIA released a blog series on Nemotron Labs showcasing how open models are enabling enterprises to build customized AI systems at a fraction of frontier model costs. The story is control: open models give teams full visibility and ownership over model behavior in a way closed models cannot, and specialized post-training can drive cost down dramatically while maintaining accuracy on domain-specific tasks.

Concrete examples from production deployments show the cost advantage. Harvey post-trained Nemotron 3 Ultra on its legal benchmark and matched frontier-class accuracy on complex legal tasks at 10x lower cost per run. Glean built Waldo, an agentic search model pairing Nemotron with larger closed models, delivering enterprise search at significantly lower latency and token cost. H Company's Holotron 3 Nano achieved 76%+ accuracy on OSWorld-Verified by post-training Nemotron 3 Nano Omni on proprietary computer-use data — matching leading frontier models at a fraction of the cost.

The broader shift is visible in infrastructure: Arcee AI achieved inference costs of roughly 90 cents per million output tokens on Nemotron post-trained on NVIDIA Blackwell — approximately 20x cheaper than comparable closed frontier alternatives while ranking second on PinchBench. For architects evaluating agent infrastructure, the trade-off is now explicit: full ownership and cost control via open models, or managed performance via closed alternatives.