Nvidia and LangChain Unveil AI Agents Ten Times Cheaper to Run

On July 8, 2026, Nvidia and LangChain unveiled a joint blueprint called NemoClaw, a ready-made set of components for building enterprise AI agents. It combines the open Nemotron 3 Ultra model, the LangChain Deep Agents Code framework, and Nvidia's OpenShell runtime environment. According to both companies, it delivers quality close to the most expensive closed models at roughly one-tenth the cost.

What the blueprint includes

NemoClaw isn't a single product but a ready-made reference architecture that engineering teams can deploy in-house and adapt to their own data and tools. Nemotron 3 Ultra handles the reasoning layer, LangChain Deep Agents Code provides the mechanism for planning and executing multi-step tasks, and Nvidia OpenShell supplies an isolated, controlled environment for running the agent's actions, such as calling external tools or performing file operations.

This architecture is meant to give companies full visibility into how an agent behaves and control over where data is physically processed, setting it apart from closed agentic services offered directly by model providers. For enterprises in regulated sectors, where data location and auditability are critical, an open architecture can be a prerequisite for deployment.

Numbers that stand out

The strongest argument in favor of NemoClaw comes from LangChain's agent evaluation benchmark. An agent built on Nemotron 3 Ultra scored 0.86 on the aggregate scale, with a total cost of $4.48 for the test run. The closest closed model in terms of quality achieved a similar score but cost $43.48 for the same run, a roughly tenfold difference.

That cost advantage matters most for agents running in the background at scale, where the number of model calls can reach thousands per day for a single company. At such volumes, the difference between a few dollars and several dozen dollars per test run translates into real savings of tens of thousands of dollars a month.

The way to build better agents is by continuously improving the system around the model - Harrison Chase, CEO and co-founder of LangChain

The era of superagents has arrived - Jensen Huang, CEO of Nvidia

A partner ecosystem

Behind the blueprint stands a broad coalition of infrastructure companies. Baseten, Fireworks, Nebius, Crusoe, DeepInfra, and Together AI offer hosting and serving of Nemotron models in production, while consulting firm EY is building a separate implementation practice around the entire stack for corporate clients. The setup is meant to make it easier for companies to choose an infrastructure provider without locking into a single cloud vendor.

The open nature of Nemotron 3 Ultra also means enterprises can train the model on their own data and run it locally, which can matter for companies in finance, healthcare, or public administration, where sending data to external cloud APIs is often restricted by regulation.

What it means for the agent market

The announcement fits a broader trend of driving down the cost of AI agents, which have often been expensive to run because of repeated model calls within a single task. Nvidia has spent months positioning its open Nemotron models as an alternative to closed APIs from OpenAI or Anthropic, betting that lower inference costs will convince companies to build their own controlled agentic stacks instead of relying on ready-made cloud services.

For Polish companies considering AI agents for process automation, this offers another, cheaper option beyond the solutions from OpenAI and Anthropic that currently dominate the market. An open model run through a chosen cloud provider also gives more flexibility in negotiating prices and choosing where data is processed, which can matter for projects subject to EU data protection rules.

Sources: LangChain Blog (langchain.com), PR Newswire (prnewswire.com), HPCwire BigDATAwire (hpcwire.com)