How are token costs calculated?

Multiply input tokens by the input rate and output tokens by the output rate, add the two amounts and divide by 1,000,000. A standard request example uses 1,000 input and 500 output tokens. Monthly budgets multiply the result by request volume; taxes, hidden reasoning tokens, tools and tier surcharges can add costs.

Which recorded model has the lowest request cost?

LFM 2.2 6b by Liquid AI costs $0.00002 for the standard request, with $0.01 input and $0.02 output per million tokens. This price ranking does not establish quality equivalence; verify task success and context requirements.

Can cached and batch discounts be combined?

Only recorded rates are used. Cached input can have cache-write or storage charges, while batch rates apply only to eligible asynchronous jobs. We do not automatically stack the discounts or treat an absent rate as zero. Check the linked provider source before budgeting.

How current is the public dataset?

The live pricing table synchronizes stored model records daily; public file snapshots update when the site is rebuilt. Synchronization is not independent provider verification. It contains 1,026 priced models across 124 providers, last updated 2026-10-31. Directory entries are counted separately. A synchronization date is not proof of independent verification.

Comparison · 2026-04-18 · 8 min read

NVIDIA Nemotron 3 Super: Pricing & Benchmarks (2026)

NVIDIA quietly entered the hosted-LLM pricing race. We benchmark Nemotron 3 Super 120B against Llama 4 and GPT-OSS on cost-per-1M-tokens, throughput, and reasoning quality.

NVIDIA Is Now an LLM Vendor, Not Just a GPU Vendor

For most of the last three years, NVIDIA's role in the LLM economy has been simple: they sell the picks and shovels. Every dollar spent on GPT-5, Claude, or Gemini eventually pays for an H100 or B200. But in late 2025 and early 2026, NVIDIA started doing something different — shipping their own family of open-weight reasoning models, Nemotron, and offering them as a hosted API on build.nvidia.com at prices that undercut almost every closed-source competitor.

Three Nemotron models matter right now for cost-conscious teams:

Nemotron 3 Super 120B — flagship reasoning model, 262K context
Llama 3.3 Nemotron Super 49B — distilled, mid-tier, 128K context
Nemotron Nano 9B v2 — small, ultra-cheap, 128K context

This post benchmarks all three against the two other big "open-ish" pricing tiers in 2026: Meta's Llama 4 family and OpenAI's GPT-OSS models — strictly on cost-per-1M-tokens for representative workloads.

Cost-Per-1M-Tokens: The Headline Numbers

All prices in USD per million tokens, current as of April 2026 on the providers' first-party hosted APIs (build.nvidia.com, api.openai.com for GPT-OSS, Together AI for Llama 4 reference pricing).

The headline: at every weight class, Nemotron is 15-40% cheaper than the equivalent Llama 4 or GPT-OSS tier on first-party pricing. Nano 9B is the cheapest serious reasoning-capable model on the market right now.

Realistic Workload: 1M Requests/Month

Headline rates are easy to misread. The number that matters is monthly cost for a real workload. Let's price out a typical RAG pipeline: 2,000 input tokens, 500 output tokens per request, 1M requests/month.

For pure cost, the picture is unambiguous — Nemotron wins every weight class. The 120B flagship is $150/month cheaper than its closest competitor (GPT-OSS 120B) and $750/month cheaper than Llama 4 Maverick at this volume.

You can run your own numbers in our calculator or compare any two of these models head-to-head.

What About Quality?

Cost only matters if quality holds up. Based on NVIDIA's published evals (MMLU-Pro, GPQA, MATH, HumanEval) and independent third-party benchmarks released in March-April 2026:

Nemotron 3 Super 120B matches or beats GPT-OSS 120B on math (MATH-500: 88.4 vs 86.1) and reasoning (GPQA: 64.2 vs 62.8). Trails Llama 4 Maverick by ~3 points on broad-knowledge MMLU-Pro but wins on tool use.
Nemotron Llama 3.3 Super 49B is essentially a distilled Llama 3.3 70B — same architecture family, ~2 points lower on most benchmarks, ~60% the cost of Llama 4 Scout.
Nemotron Nano 9B v2 punches above its weight: HumanEval 76.8 (better than Llama 3.1 8B's 66.5) at the cheapest hosted price of any 2026 reasoning model.

The summary: for math, code, and tool-use heavy workloads, Nemotron is now the price-performance leader. For broad-knowledge chat and creative writing, Llama 4 Maverick still has an edge — but at 75% higher cost.

Throughput and Latency

NVIDIA hosts Nemotron on their own infrastructure, which means it runs on the latest Blackwell B200 silicon. Reported tokens/second on build.nvidia.com:

Nano 9B is the throughput champion of the group — useful for streaming UIs and real-time agents where time-to-first-token matters more than per-token quality.

Caveats Before You Migrate

A few things to know before you swap your production traffic:

No batch API yet. OpenAI and Anthropic both offer 50% discounts via Batch API. NVIDIA does not (as of April 2026). For async-heavy workloads, GPT-OSS via OpenAI's Batch API can actually be cheaper than Nemotron — see our Batch API calculator.
Rate limits are tighter than OpenAI/Anthropic on the free tier. Pay-as-you-go tier matches industry norms.
No prompt caching discount — Anthropic and OpenAI both offer ~90% off cached input tokens. For long-system-prompt workloads, this can flip the cost equation.
Open weights = self-hosting option. All three Nemotron models are released under permissive licenses. If your monthly bill exceeds ~$5,000, self-hosting on rented H100s starts to win.

Who Should Switch

High-volume RAG and document-processing teams running on GPT-OSS 120B → Nemotron 3 Super 120B saves ~13% with similar quality.
Cost-sensitive startups running cheap classification or extraction → Nemotron Nano 9B v2 is the cheapest serious model on the market.
Llama 4 Maverick users who don't need broad-knowledge reasoning → Nemotron 3 Super 120B saves ~43% on the same workload.

Who Should Stay Put

Heavy Batch API users on OpenAI or Anthropic — the 50% async discount still wins.
Long-system-prompt workloads with prompt caching active on Anthropic — the 90% cache discount is hard to beat.
Anything needing GPT-5-class broad reasoning — Nemotron isn't there yet.

Bottom Line

NVIDIA Nemotron is the most cost-effective hosted reasoning family on the market in April 2026, full stop. If you're paying for GPT-OSS or Llama 4 today and your workload is reasoning-heavy or budget-constrained, the migration math almost always favors switching.

Run your own numbers — Nemotron 3 Super 120B, Llama 3.3 Nemotron Super 49B, Nemotron Nano 9B v2 — or compare any pair on the model comparison tool.