NVIDIA Nemotron 3 Ultra 550B A55B is a 550B-parameter mixture of experts with 55B active per token, NVIDIA's open-weights flagship. Inside Atlas it is the frontier-scale build model, and the interesting thing is how differently it is priced: the same weights range from $0.50 to $0.60 input and $2.20 to $3.60 output per Mtok depending on who is serving them. NVIDIA NIM is $0.50/$2.50 per Mtok with a 1,000,000 token window. OpenRouter is $0.50/$2.20. Together AI is $0.60/$3.60 and caps context at 512,300 tokens.
What is NVIDIA Nemotron 3 Ultra 550B A55B best at inside Atlas?
NVIDIA Nemotron 3 Ultra 550B A55B is best inside Atlas at hard, multi-file reasoning. With 550B total parameters and only 55B active per token, you get frontier-scale capacity at roughly a tenth of the per-token compute of a dense model that size.
Nemotron 3 Ultra 550B A55B earns the build-agent slot when a change spans many files and the model must hold invariants that no smaller model retains. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, which is the correct shape for a model of this class: let Nemotron 3 Ultra reason through the approach, read the plan, then authorize the write. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so a 550B model's confidence never translates into an unreviewed change on disk. Atlas snapshots file changes as git patches so edits can be diffed and rolled back.
Why does the same Nemotron 3 Ultra cost different amounts on different hosts?
NVIDIA Nemotron 3 Ultra 550B A55B ships open-weights, so five services resell byte-identical weights at five prices. The output price spread across hosts is 64 percent, OpenRouter at $2.20 per Mtok against Together AI at $3.60, for exactly the same model.
Nemotron 3 Ultra's 2026 listings read: NVIDIA NIM $0.50/$2.50 per Mtok, OpenRouter $0.50/$2.20, Vercel AI Gateway $0.60/$2.40, Baseten $0.60/$2.40, Together AI $0.60/$3.60. Provider choice alone changes your bill materially, and because the weights are identical there is no quality argument to weigh against the price difference, only context. Pick your host by price, not by name. In Atlas the host is part of the model string in atlas.json, so moving Nemotron 3 Ultra off Together AI's $3.60 output and onto OpenRouter's $2.20 is a single line edit that cuts generation cost by more than a third.
Is the 1,000,000 token context on Nemotron 3 Ultra real?
The advertised million-token context is only real on some hosts. NVIDIA Nemotron 3 Ultra 550B A55B gets 1,000,000 tokens on NVIDIA NIM and Vercel AI Gateway, 512,300 on Together AI, and 262,144 on Ollama Cloud and Baseten, a quarter of the headline number.
Context on Nemotron 3 Ultra is a serving decision, and the gap between hosts is large enough to change what you can attempt. The 1,000,000 token context on NVIDIA NIM and Vercel is 4x the 262,144 window the same model gets on Ollama Cloud and Baseten. Pinning `togetherai/nvidia/nemotron-3-ultra-550b-a55b` gets you 512,300 tokens, not 1M. Verify before you trust it: run `atlas models nvidia` and confirm the `nvidia/nemotron-3-ultra-550b-a55b` row shows the context you expect. What fills that window matters as much as its size. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and Atlas indexes code by AST declarations using tree-sitter, not blind line windows.
Why is Nemotron 3 Ultra expensive to run as a coding agent?
Reasoning is on by default on NVIDIA Nemotron 3 Ultra 550B A55B, so it burns output tokens on thinking traces, and output is the expensive side at $2.20 to $3.60 per Mtok against $0.50 to $0.60 input. Thinking, not context, is what lands on the invoice.
Nemotron 3 Ultra inverts the usual instinct to trim context. Feeding it a deep repo slice at $0.50 per Mtok on NVIDIA NIM is inexpensive. Letting it think at length at $2.50 per Mtok of output is not. The defensive move is to keep the 550B model away from work that does not need it. Pair it with a cheap `small_model` (Nemotron 3 Nano at $0.05/$0.20 per Mtok on DeepInfra) so Atlas's title and summary calls do not run through a 550B model. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and every one of those subagents is another opportunity to spend $2.50 per Mtok that a Nano-class model would have covered.
When should you use something other than Nemotron 3 Ultra?
Use something other than NVIDIA Nemotron 3 Ultra 550B A55B for any work that does not need frontier reasoning. At $2.20 to $3.60 per Mtok output with reasoning on by default, a 550B mixture of experts is the most expensive possible way to draft a commit message.
Nemotron 3 Ultra is the wrong tool for routine overhead and the right tool for genuinely hard problems, and in Atlas that distinction is a two-slot config rather than a compromise. Keep `"model"` on Nemotron 3 Ultra and put a Nano-class model in `"small_model"`. Because a large model produces large changes, the guardrails matter more here than anywhere else: every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, so a 550B model operating on your repository still asks before it acts.
Setup
- 01Pick your host by price, not by name. Export NVIDIA_API_KEY for the $0.50/$2.50 NIM listing at the full 1M context.
- 02Run `atlas models nvidia` and confirm the `nvidia/nemotron-3-ultra-550b-a55b` row shows the context you expect.
- 03Pin `"model": "nvidia/nvidia/nemotron-3-ultra-550b-a55b"` in atlas.json.
- 04Pair it with a cheap `small_model` (Nemotron 3 Nano at $0.05/$0.20 per Mtok on DeepInfra) so Atlas's title and summary calls do not run through a 550B model.
- 05If the million-token window is not required, compare the OpenRouter listing at $0.50/$2.20 per Mtok, the cheapest output price for these weights.
Frequently asked questions
- how much does nvidia nemotron 3 ultra cost per million tokens
- Nemotron 3 Ultra 550B A55B is $0.50/$2.50 per Mtok on NVIDIA NIM, $0.50/$2.20 on OpenRouter, $0.60/$2.40 on Vercel AI Gateway and Baseten, and $0.60/$3.60 on Together AI.
- does nemotron 3 ultra really have a 1 million token context
- Only on NVIDIA NIM and Vercel AI Gateway. Together AI serves Nemotron 3 Ultra at 512,300 tokens, and Ollama Cloud and Baseten serve it at 262,144, a quarter of the advertised window.
- how many parameters does nemotron 3 ultra 550b a55b have
- Nemotron 3 Ultra is a 550B-parameter mixture of experts with 55B active per token, NVIDIA's open-weights flagship. Only 55B activate on any token, giving frontier-scale capacity at roughly a tenth of a dense 550B model's per-token compute.
- which provider is cheapest for nemotron 3 ultra
- OpenRouter has the cheapest output at $0.50/$2.20 per Mtok. NVIDIA NIM at $0.50/$2.50 is the cheapest listing that still gives the full 1,000,000 token context, so the answer depends on whether you need the window.
- how do I run atlas on nemotron 3 ultra
- Export NVIDIA_API_KEY, run `atlas models nvidia` and confirm the `nvidia/nemotron-3-ultra-550b-a55b` row, then pin `"model": "nvidia/nvidia/nemotron-3-ultra-550b-a55b"` in atlas.json.
- why is nemotron 3 ultra expensive for agentic coding
- Reasoning is on by default, so Nemotron 3 Ultra burns output tokens on thinking traces, and output is the expensive side at $2.20 to $3.60 per Mtok against $0.50 to $0.60 input.
- what small model pairs well with nemotron 3 ultra
- Nemotron 3 Nano at $0.05/$0.20 per Mtok on DeepInfra. Pair it as your `small_model` so Atlas's title and summary calls do not run through a 550B model.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Crystal: A Terminal-Native AI Coding Agent for shard.yml Projects in 2026
Atlas is a terminal-native AI coding agent for Crystal in 2026. It reads shard.yml and union types, narrows Nil unions, runs crystal spec behind a prompt, and formats the diff.
Atlas for PyTorch: Terminal-Native AI Coding for nn.Module, Devices, and Autograd in 2026
Atlas is a terminal-native AI coding agent for PyTorch in 2026, where device placement, autograd, and DataLoader worker counts cause most bugs and most slowness.
Atlas for Bash in 2026
Atlas is a terminal-native AI coding agent for Bash in 2026, with shellcheck-aware suggestions. Harden a script or fix shellcheck warnings and review every diff.
Rename a Symbol Across the Repo with Atlas in 2026
How to rename a symbol across a repo with Atlas in 2026: findReferences gets the true reference set, grep catches strings and docs, and edit refuses ambiguous matches.
Atlas for R: A Terminal-Native AI Coding Agent for tidyverse, roxygen2, and testthat in 2026
Atlas is a terminal-native AI coding agent for R in 2026. It reads roxygen2 docblocks and renv.lock, rewrites loops as dplyr or purrr pipelines, and runs devtools::test().
Atlas vs Poolside: Terminal AI Coding Agents in 2026
Compare Atlas and Poolside, two terminal AI coding agents in 2026. Atlas offers a free core and robust safety features, while Poolside provides open-weight Laguna models and on-prem deployment options.
Atlas for SQL in 2026
Atlas is a terminal-native AI coding agent for SQL in 2026. Run it in a repo with your migrations or .sql files, optimize a query, and review the diff before applying.
Atlas for FastAPI in 2026
Atlas is a terminal-native AI coding agent for FastAPI in 2026. It reads Pydantic v2 models and Depends(), then runs pytest behind a permission prompt.