NVIDIA Nemotron 3 Super 120B A12B is the 120B/12B-active middle of the Nemotron 3 line, and it is the sweet spot: reasoning-capable, cheap, and available on more hosts than any other Nemotron. Vercel AI Gateway lists it at $0.15/$0.65 per Mtok, NVIDIA NIM at $0.20/$0.80, Baseten at $0.30/$0.75, Nebius at $0.30/$0.90, and Cloudflare Workers AI at $0.50/$1.50. NVIDIA NIM carries 262,144 tokens of context with 262,144 max output. The price spread across those hosts is wider than the spread between entire model tiers.
Which host is cheapest for NVIDIA Nemotron 3 Super 120B A12B?
Vercel AI Gateway is the cheapest host for Nemotron 3 Super 120B A12B at $0.15/$0.65 per Mtok. Cloudflare Workers AI is the most expensive at $0.50/$1.50. Input price varies 3.3x across hosts for identical weights, so where you point Atlas is a real cost decision.
Five hosts carry Nemotron 3 Super 120B A12B in 2026: Vercel AI Gateway $0.15/$0.65 per Mtok, NVIDIA NIM $0.20/$0.80, Baseten $0.30/$0.75, Nebius $0.30/$0.90, Cloudflare Workers AI $0.50/$1.50. The same model is a different product depending on where you point Atlas. Cloudflare's listing is both the most expensive ($0.50/$1.50 per Mtok) and no larger in context than the cheapest, so there is no upside to that route. The default recommendation is to pin `"model": "vercel/nvidia/nemotron-3-super-120b-a12b"` in atlas.json at $0.15/$0.65 per Mtok, which is the best price on these weights anywhere.
Why is a 120B reasoning model priced at $0.15 per Mtok?
Nemotron 3 Super has 120B total parameters with 12B active per token, which is why it lands at $0.15 to $0.20 input per Mtok despite being a reasoning model. Serving cost tracks the 12B that fire, not the 120B that exist.
Sparse activation is what makes Nemotron 3 Super's pricing possible. Only 12B of the 120B parameters engage on any given token, so a host can quote $0.15 per Mtok on Vercel AI Gateway for a model whose capacity is far larger than a 12B dense model's. For Atlas users, the effect is that a reasoning-grade model costs roughly what a small non-reasoning model costs from other vendors. Atlas is extensible through plugins that contribute tools and hook into agent lifecycle events, and Atlas connects to Model Context Protocol servers and exposes their tools to the agent, so a cheap reasoning model can be handed real tools rather than being limited to text generation.
What is the max output on Nemotron 3 Super, and why does it matter?
NVIDIA NIM serves Nemotron 3 Super with max output equal to its full 262,144 token context, so a long reasoning trace plus a large diff fits in one call. Vercel, Nebius, and Cloudflare Workers AI serve the model at 256,000 tokens instead.
Max output is the one dimension where paying NVIDIA NIM's $0.20/$0.80 per Mtok over Vercel's $0.15/$0.65 buys you something concrete. A reasoning model spends output tokens on thinking before it spends them on the answer, and a low output ceiling means the thinking crowds out the patch. NIM's 262,144 token output ceiling removes that competition entirely. Atlas reads git branches, status, and diffs, and can stage and create commits on your behalf, so a model that can emit a large change in one call fits naturally into a commit-shaped workflow rather than forcing artificial chunking.
What does Nemotron 3 Super actually cost you per session?
Reasoning traces on Nemotron 3 Super consume output tokens, and output is priced 3x to 4x input on every host, from $0.65 per Mtok on Vercel AI Gateway to $1.50 on Cloudflare Workers AI. A chatty thinking budget is what actually costs you money.
Nemotron 3 Super's input price is low enough on Vercel AI Gateway at $0.15 per Mtok that filling a 256,000 token window is not the line item that hurts. Generation is. Every one of the five hosts marks output up 3x to 4x over input, so session cost tracks how much the model thinks, not how much you show it. Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers, which means the indexing side of a large repo does not add to your Nemotron 3 Super bill at all. Only the agent turns do.
How do you flip between Nemotron 3 Super and a frontier model mid-session?
Add Nemotron 3 Super to favorites in the Atlas `/models` dialog so `model.cycle_recent` can flip between Super and a frontier model mid-session. Nemotron 3 Super at $0.15/$0.65 per Mtok handles most turns; a frontier model handles the ones it cannot.
Nemotron 3 Super is a strong default rather than a universal answer, and the right pattern is to keep an escape hatch bound to a keystroke. Atlas lets you switch the active model and provider on the fly with favorites and recents, so the moment a task turns out to need more than 12B active parameters, you move without restarting the session. Atlas ships a TUI theme system with a charcoal-and-blue default theme and many presets, and Atlas is a terminal-native TUI rendered with SolidJS through the OpenTUI renderer, so the model switcher is part of the same terminal surface you are already working in.
Setup
- 01Export AI_GATEWAY_API_KEY for the cheapest listing, or NVIDIA_API_KEY for the largest output window.
- 02Run `atlas models vercel` (or `atlas models nvidia`) and confirm the nemotron-3-super row and its context.
- 03Pin `"model": "vercel/nvidia/nemotron-3-super-120b-a12b"` in atlas.json at $0.15/$0.65 per Mtok.
- 04Add it to favorites in the `/models` dialog so `model.cycle_recent` can flip between Super and a frontier model mid-session.
- 05Skip the Cloudflare Workers AI listing at $0.50/$1.50 per Mtok, which costs more without giving you a larger context.
Frequently asked questions
- how much does nemotron 3 super 120b cost per million tokens
- Nemotron 3 Super 120B A12B is $0.15/$0.65 per Mtok on Vercel AI Gateway, $0.20/$0.80 on NVIDIA NIM, $0.30/$0.75 on Baseten, $0.30/$0.90 on Nebius, and $0.50/$1.50 on Cloudflare Workers AI.
- what is the context window of nemotron 3 super
- Nemotron 3 Super holds 262,144 tokens on NVIDIA NIM, with max output equal to that same 262,144 token context. Vercel AI Gateway, Nebius, and Cloudflare Workers AI serve it at 256,000 tokens.
- which provider is cheapest for nemotron 3 super
- Vercel AI Gateway at $0.15/$0.65 per Mtok. Cloudflare Workers AI at $0.50/$1.50 is the most expensive and offers no larger context, so there is no upside to that route.
- how do I configure nemotron 3 super in atlas
- Export AI_GATEWAY_API_KEY, run `atlas models vercel` and confirm the nemotron-3-super row, then pin `"model": "vercel/nvidia/nemotron-3-super-120b-a12b"` in atlas.json at $0.15/$0.65 per Mtok.
- is nemotron 3 super a reasoning model
- Yes. Nemotron 3 Super 120B A12B is reasoning-capable, and its reasoning traces consume output tokens, which every host prices at 3x to 4x input. That thinking budget is what actually costs you money.
- nemotron 3 super vs nemotron 3 ultra for coding
- Nemotron 3 Super is 120B with 12B active and starts at $0.15/$0.65 per Mtok. Nemotron 3 Ultra is 550B with 55B active and costs $0.50 to $0.60 input with $2.20 to $3.60 output. Super is the cheap reasoning default.
- can nemotron 3 super emit a large diff in one call
- On NVIDIA NIM, yes. NVIDIA NIM serves Nemotron 3 Super with max output equal to its full 262,144 token context, so a long reasoning trace plus a large diff fits in a single call.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Fiber in 2026
Atlas is a terminal-native AI coding agent for Fiber in 2026. It knows fasthttp reuses buffers, tests handlers with app.Test(), and diffs every edit first.
Atlas for React Native: Terminal-Native AI Coding Across the Native Boundary in 2026
Atlas is a terminal-native AI coding agent for React Native in 2026. Work across the New Architecture, native modules, and platform-specific files with diff-first review.
Atlas vs Mistral Vibe for Code: Terminal AI Coding Agents in 2026
Compare Atlas and Mistral Vibe for Code in 2026. Atlas offers terminal-native TUI, explicit diffs, and BYO models. Mistral Vibe for Code provides a four-model stack, multi-platform access, and EU data sovereignty.
Atlas for Axum in 2026
Atlas is a terminal-native AI coding agent for Axum in 2026. It decodes tower trait-bound errors, adds IntoResponse types, and runs cargo nextest run.
Atlas for Angular in 2026
Adopt Atlas, the terminal-native AI coding agent, for your Angular projects in 2026. Enhance development with intelligent code search, secure local embeddings, and granular control over AI actions.
Atlas vs Cosine: Terminal AI Coding Agents in 2026
Comparing Atlas and Cosine for terminal AI coding in 2026. Atlas offers a free core and local privacy, while Cosine provides proprietary models and a cloud surface.
Atlas for Nim: A Terminal-Native AI Coding Agent for Nimble Packages and Macros in 2026
Atlas is a terminal-native AI coding agent for Nim in 2026. It reads .nimble requires and asterisk-exported symbols, adds std/unittest suites, runs nimble test, formats with nph.
Atlas vs GitHub Copilot: Terminal AI Coding Agents in 2026
Comparing Atlas, the terminal-native AI coding agent, with GitHub Copilot's editor extension and chat features for developers in 2026. Explore planning, pricing, and privacy.