# Nvidia Nemotron 3 Ultra: 10x Cheaper Than Closed

> Source: [https://icharles.com/articles/nvidia-nemotron-ultra-open-model-benchmark](https://icharles.com/articles/nvidia-nemotron-ultra-open-model-benchmark) (canonical)
> Author: iCharles News — iCharles, https://icharles.com
> Published: 2026-07-12

## TL;DR

Nvidia's Nemotron 3 Ultra, a 550-billion-parameter open model, matched the highest-scoring closed AI models on LangChain's Deep Agents benchmark while delivering inference costs 10 times lower per run. The news sent Nvidia shares up 4% on Wednesday. Enterprise customers including Abridge, Amdocs, Box, and EY are already embedding the model into their platforms. The model is available free via OpenRouter and carries a 1-million-token context window.

## What is Nvidia Nemotron 3 Ultra?

**Nemotron 3 Ultra** is an open frontier-reasoning model from Nvidia with 550 billion total parameters and 55 billion active parameters. It uses a hybrid Transformer-Mamba mixture-of-experts (MoE) architecture. It supports text input and output with a context window of up to 1 million tokens, according to [OpenRouter's model listing](https://openrouter.ai/nvidia/nemotron-3-ultra-550b-a55b:free).

The model was released on June 4, 2026. Nvidia positions it for long-running agentic workflows: agent orchestration, coding agents, deep research, and complex enterprise tasks.

## How did Nemotron 3 Ultra perform on the LangChain benchmark?

Nvidia reported that LangChain's Deep Agents harness completed more tasks at higher throughput using Nemotron 3 Ultra. Inference costs per run came in at **10 times lower** than some leading closed models, [according to GuruFocus via TradingView](https://www.tradingview.com/news/gurufocus:6f69731b8094b:0-nvidia-shares-rise-as-nemotron-3-ultra-challenges-closed-ai-models/).

On the same benchmark, Nvidia said the model reached business-task parity with the highest-scoring models. That means it matched closed competitors on the quality of enterprise task completion while running at a fraction of the cost.

LangChain CEO Harrison Chase said enterprises can achieve strong performance with an open stack while retaining control over the agent systems they build.

## Why did Nvidia stock rise 4%?

Nvidia shares gained 4% on Wednesday after the benchmark results were published. The results suggested an open-source model could challenge closed AI on both cost and enterprise performance — a meaningful signal for buyers evaluating [Nvidia AI infrastructure](/articles/hesai-nvidia-lidar-security-risk) decisions.

Here's what we know so far: the stock move tracked directly to the benchmark release and the enterprise adoption announcements, with no other major Nvidia news reported on the same day.

## Which enterprises are already using Nemotron 3 Ultra?

Nvidia pointed to four named enterprise customers embedding the model or related capabilities:

- **Abridge** — embedding specialized agents into its platform
- **Amdocs** — embedding specialized agents into its platform
- **Box** — embedding specialized agents into its platform
- **EY** — expanding use of Nvidia capabilities around NemoClaw blueprints

The NemoClaw blueprints are a separate Nvidia offering that EY is scaling alongside Nemotron 3 Ultra adoption.

## What are Nemotron 3 Ultra's technical specs?

| Spec | Detail |
|---|---|
| Total parameters | 550B |
| Active parameters | 55B |
| Architecture | Hybrid Transformer-Mamba MoE |
| Context window | 1M tokens |
| Release date | June 4, 2026 |
| Modalities | Text in, text out |
| OpenRouter price | Free |

The model is part of Nvidia's Nemotron family of open models for [agentic AI](/articles/nvidia-kyber-nvl144-delay-2028). It is designed for high-throughput inference in high-volume agent pipelines, per [OpenRouter's listing](https://openrouter.ai/nvidia/nemotron-3-ultra-550b-a55b:free).

## How does Nemotron 3 Ultra compare to closed AI models on cost?

Nvidia's claim is specific: inference costs per run on LangChain's Deep Agents harness were 10 times lower than some leading closed models. The model also completed more tasks at higher throughput in the same benchmark run.

On OpenRouter, the model currently carries a $0.00 per million token price for both input and output via Nvidia's free endpoint. The provider shows 47 tokens per second throughput and a 2.03-second latency at the p50 level.

## Where can developers access Nemotron 3 Ultra today?

Developers can call the model through OpenRouter using the slug `nvidia/nemotron-3-ultra-550b-a55b:free`. OpenRouter's API is OpenAI-compatible, so most existing SDKs work by swapping the base URL.

The free endpoint logs session data for security and product improvement purposes. Nvidia states that logged data is not linked to user identity or any persistent identifier.

OpenRouter data shows the model is already seeing significant production traffic. The top apps sending traffic to it include Hermes Agent (830 billion tokens), Claude Code (72.2 billion tokens), and Kilo Code (71 billion tokens).

## What is the broader context for open vs. closed AI models?

Nvidia's push with Nemotron 3 Ultra fits a wider pattern of open models closing the gap on closed ones. The [Nvidia Kyber NVL144 hardware roadmap](/articles/nvidia-kyber-nvl144-delay-rubin-ultra) and ongoing chip investments are part of the same infrastructure build-out that makes running large open models viable at enterprise scale.

The LangChain benchmark result — business-task parity with the highest-scoring closed models — is the most concrete data point Nvidia has published to support that claim. EY's expansion around NemoClaw blueprints and Box's agent embedding are the named enterprise proof points backing the adoption narrative.

## Frequently asked questions

**What is Nvidia Nemotron 3 Ultra?**

Nvidia Nemotron 3 Ultra is an open AI model with 550 billion total parameters and 55 billion active parameters. It uses a hybrid Transformer-Mamba mixture-of-experts architecture, supports a 1-million-token context window, and was released on June 4, 2026. It is designed for agentic workflows including coding agents, deep research, and complex enterprise tasks.

**How much cheaper is Nemotron 3 Ultra than closed AI models?**

Nvidia reported that inference costs per run on LangChain's Deep Agents harness were 10 times lower than some leading closed models. On OpenRouter, the model is currently available at no cost — $0.00 per million tokens for both input and output via Nvidia's free endpoint, with 47 tokens per second throughput.

**What benchmark did Nemotron 3 Ultra use to compare against closed models?**

Nvidia used LangChain's Deep Agents benchmark. On that evaluation, Nemotron 3 Ultra completed more tasks at higher throughput than competing runs, and reached business-task parity with the highest-scoring models. LangChain CEO Harrison Chase said enterprises can achieve strong performance with an open stack while retaining control over their agent systems.

**Which companies are using Nvidia Nemotron 3 Ultra in production?**

Nvidia named four enterprise customers: Abridge, Amdocs, and Box are embedding specialized agents into their platforms using the model. EY is expanding its use of Nvidia capabilities around NemoClaw blueprints. These were cited by Nvidia as evidence of growing enterprise adoption alongside the benchmark announcement.

**How can developers access Nemotron 3 Ultra for free?**

Developers can access the model through OpenRouter using the model slug `nvidia/nemotron-3-ultra-550b-a55b:free`. OpenRouter's API is OpenAI-compatible. Nvidia's free endpoint logs session data for security and product improvement but states it is not linked to user identity or any persistent identifier.