Loading page
NVIDIA
Nemotron 3 Ultra 550B A55B
Nemotron 3 Ultra 550B A55B is a large-scale model from NVIDIA in its Nemotron 3 family, positioned at the Ultra tier with roughly 550B total and 55B active parameters per token.
Input from
$0.600 / 1M tokens
across 4 providers
API Pricing
Cheapest on Together AI — 20% below avg| Provider | Input / 1M | Output / 1M | Cached / 1M |
|---|---|---|---|
| $0.600 | $3.60 | - | |
| $0.600 | $2.40 | $0.120 | |
| $0.800 | $2.60 | $0.100 | |
| $1.00 | $3.20 | $0.250 |
Prices updated daily. Last check: Sep 22, 2026
Nemotron 3 Ultra 550B A55B pricing by provider
Input, output, and batch rates, plus alternatives, for one provider at a time.
Performance & Benchmarks
Source: Artificial Analysis →Intelligence
22.9 / 100
Coding
49.3 / 100
Output Speed
173 t/s
Latency (TTFT)
1.2s
Reasoning & Knowledge
- GPQA Diamond86.7%
- Humanity's Last Exam28.4%
Coding
- SciCode40.3%
Agentic & Tool Use
- Terminal-Bench Hard36.4%
- Terminal-Bench v2.153.9%
- τ²-bench83.3%
- τ-bench Banking14.2%
Instruction & Long Context
- IFBench81.4%
- Long-Context Reasoning79.3%
Benchmarks measured Sep 2026. Scores are independent evaluations, not vendor-reported.
Model Details
General
- Creator
- NVIDIA
- Modalities
- Text
Capabilities
- Tool Calling
- No
- Open Source
- No
Strengths & Limitations
Strengths
- Sparse mixture-of-experts naming (550B total / 55B active) implies substantially lower per-token compute than a dense model of comparable total size
- Measured output speed of roughly 139 tokens per second in Artificial Analysis testing, which is competitive for a model of this scale
- Ultra tier sits at the top of NVIDIA's Nemotron 3 sizing ladder, giving teams a same-family upgrade path from smaller Nemotron 3 variants
- Part of NVIDIA's Nemotron program, which is closely aligned with NVIDIA's own inference software stack and GPU deployment tooling
- Time to first token of about 1.35 seconds measured by Artificial Analysis, workable for streaming chat and agent loops
- Large-scale capacity suits demanding workloads such as synthetic data generation and multi-step reasoning pipelines
Limitations
- We do not currently track a confirmed context window for this model — verify the limit with NVIDIA or your chosen provider before building around long inputs
- No standardized benchmark scores (reasoning, coding, math) are recorded in our database for this model, making direct capability comparisons harder
- Supported input modalities are not confirmed in our data; do not assume image or audio input without checking the model card
- A 550B-parameter total footprint makes self-hosting impractical without substantial multi-GPU infrastructure
- Time to first token above one second is slower than small, latency-tuned models used for interactive autocomplete or classification
Key Features
•Mixture-of-experts architecture implied by the 550B total / 55B active parameter naming
•Ultra tier placement within NVIDIA's Nemotron 3 model family
•Measured throughput of ~139 output tokens per second (Artificial Analysis)
•Measured time to first token of ~1347 ms (Artificial Analysis)
•Streaming text generation via hosted inference endpoints
•Alignment with NVIDIA's Nemotron model program and GPU inference tooling
•Multiple Nemotron 3 size tiers available for routing work by difficulty
About Nemotron 3 Ultra 550B A55B
Nemotron 3 Ultra 550B A55B is a large language model released by NVIDIA as part of the Nemotron 3 line. The "Ultra" designation places it at the upper end of the Nemotron 3 sizing ladder, above the smaller members of the family. The "550B A55B" portion of the name follows NVIDIA's convention for sparse mixture-of-experts models, indicating approximately 550 billion total parameters with about 55 billion active per token — a design that keeps per-token compute far below what a dense model of the same total size would require.
Our database records third-party performance measurements for this model rather than a full capability profile. Artificial Analysis measured output generation at roughly 139 tokens per second with a time to first token of about 1.35 seconds, which puts it in the range of a large model served on optimized inference stacks rather than a small, latency-optimized endpoint. We do not currently track a confirmed context window, modality list, or benchmark suite scores for Nemotron 3 Ultra 550B A55B, so buyers evaluating it for a specific workload should confirm those details against NVIDIA's own model card and the documentation of whichever provider they intend to use.
In practice, models at this scale and tier are usually deployed for the harder end of a workload mix — multi-step reasoning, code generation and review, synthetic data generation, and agentic pipelines — while smaller family members handle high-volume, low-complexity traffic. Because Nemotron models come from NVIDIA rather than a closed API-first lab, availability and serving characteristics can vary meaningfully between hosts, so the pricing table on this page is the practical starting point for comparing providers.
Common Use Cases
Nemotron 3 Ultra 550B A55B is aimed at the workloads where teams reach for the largest model in a family: multi-step reasoning chains, code generation and refactoring across larger contexts, document analysis, and agentic systems where an error early in a chain compounds. Its measured throughput of roughly 139 tokens per second makes it usable for streaming assistant interfaces and background batch jobs alike, though the ~1.3 second time to first token makes it a poor fit for ultra-low-latency tasks like inline autocomplete or high-volume classification — those are better served by smaller Nemotron 3 tiers or other lightweight models. Because the Ultra tier represents the top of NVIDIA's Nemotron 3 range, a common pattern is to route only the hardest requests here and handle the rest with cheaper models, comparing per-provider costs in the pricing table on this page to decide where that cutoff should fall.
Frequently Asked Questions
How much does Nemotron 3 Ultra 550B A55B cost to use?
Pricing varies by provider and by pricing type — hosted token-based APIs, batch or cached-input rates, and self-managed GPU deployments all price differently, and rates change frequently. Check the pricing table on this page for current per-provider figures rather than relying on a fixed number.
What is Nemotron 3 Ultra 550B A55B best used for?
It is best suited to the harder end of a workload mix: multi-step reasoning, code generation and review, document analysis, synthetic data generation, and agentic pipelines where accuracy at each step matters. Lower-complexity, high-volume traffic is usually more economical on smaller Nemotron 3 tiers.
What does "550B A55B" mean in the name?
It follows NVIDIA's naming convention for sparse mixture-of-experts models: roughly 550 billion total parameters with about 55 billion active for any given token. The practical effect is that inference compute scales with the active parameter count, not the full total, while the model still draws on the larger parameter pool.
How fast is Nemotron 3 Ultra 550B A55B?
Artificial Analysis measured approximately 139 output tokens per second with a time to first token of about 1347 ms. Actual figures depend heavily on the provider, hardware, batch size, and prompt length, so treat these as a reference point rather than a guarantee.
What context window does it support?
We do not have a confirmed context window recorded for Nemotron 3 Ultra 550B A55B. Check NVIDIA's model card and your chosen provider's documentation before designing around long-context inputs, since hosts sometimes serve a shorter window than the model supports.
How does it compare to other Nemotron 3 models?
The Ultra designation places it at the top of the Nemotron 3 sizing range, so it is intended for tasks that smaller family members struggle with. The usual trade-off applies: more capacity per request, but higher cost per token and slower first-token latency than the smaller tiers.