Skip to content
CostPerPrompt

Nemotron 3 Ultra (batch) API Pricing

NVIDIA · context window 512K · prices updated 2026-08-24

Input per 1M tokens
$0.6
Output per 1M tokens
$3.60
Cached input per 1M
$0.2
67% cheaper than fresh input

Where Nemotron 3 Ultra (batch) sits on price

At $3.60 per million output tokens, Nemotron 3 Ultra (batch) is a mid tier model — 1.6× the median output price of $2.25, which puts it cheaper than 41% and pricier than 59% of the models we track. It is the most expensive of the 5 NVIDIA models we track. Output costs 6.0× more than input here, so anything that makes the model write less — tighter instructions, structured output, lower max_tokens — moves the bill more than trimming the prompt.

Output tokens per $1
277,778
One full context fill
$0.3074
Cheaper than
41% of tracked models

Cached input is 67% cheaper. On an input-heavy workload you need roughly a 37% cache-hit rate to take 25% off the input line — reachable for chatbots and agents that resend the same system prompt and history.

What real workloads cost on Nemotron 3 Ultra (batch)

These three workloads are the ones teams actually run on a mid-priced model — a $3.60/1M model is not bought for the same job as one ten times the price.

Workload Per request Per month
Support chatbot — 500 conversations/day (5K in / 1.4K out) Conversation history re-sent each turn — the classic prompt-caching win. $0.008 $121
Document pipeline — 10K docs/day (3K in / 500 out) Batch-friendly: no user waiting, so a ~50% batch discount usually applies. $0.0036 $1080
Long-context analysis — 200 runs/day (250K in / 4K out) Only possible on large-context models; input cost is nearly the whole bill. $0.1644 $986

Model your exact traffic in the API cost calculator — it preloads Nemotron 3 Ultra (batch) with caching and batch options.

Nemotron 3 Ultra (batch) price history

NVIDIA last raised Nemotron 3 Ultra (batch)'s output price on Aug 21, 2026 (+100% per output token), from $0.3/$1.80 to $0.6/$3.60 per 1M in/out. That is the only repricing in the 18 days we have tracked it — worth knowing before you hard-code today's rate into a budget.

Date Input /1M Output /1M Output change
Aug 6, 2026 (tracking began) $0.3 $1.80
Aug 21, 2026 $0.6 $3.60 +100%

Change-points from our daily price snapshots (tracking since Aug 6, 2026; intraday moves between snapshots are not captured).

Nemotron 3 Ultra (batch) vs Qwen3.5 397B A17B

The closest-priced alternative from another vendor is Qwen3.5 397B A17B (Alibaba (Qwen)) — priced within a rounding error on output, with $0.1 less per million input tokens. When two models land this close on price, the decision is quality on your own workload, not the price sheet: run 50 real requests through both and compare.

See Qwen3.5 397B A17B pricing →

Cheaper alternatives

More NVIDIA models

Frequently asked questions

How much does the Nemotron 3 Ultra (batch) API cost?

Nemotron 3 Ultra (batch) costs $0.6 per million input tokens and $3.60 per million output tokens, with cached input at $0.2 per million (67% cheaper). That works out to roughly 277,778 output tokens per dollar.

What does the support chatbot workload cost on Nemotron 3 Ultra (batch)?

Support chatbot — 500 conversations/day (5K in / 1.4K out) costs about $0.008 per request and $121 per month on Nemotron 3 Ultra (batch). Conversation history re-sent each turn — the classic prompt-caching win.

What does it cost to fill Nemotron 3 Ultra (batch)'s 512K context window?

Sending 512K of input in a single request costs $0.3074 at $0.6 per million tokens — before any output. With prompt caching that same fill drops to about $0.1025 on repeat requests. This is why large context windows are cheap to advertise and expensive to actually use.

Is Nemotron 3 Ultra (batch) worth the price?

Nemotron 3 Ultra (batch) sits in the middle of the market (5 of 5 by price within NVIDIA). The honest test is a routing experiment: send the same 200 real requests to this model and to a tier below, and compare failure rate against the price gap — most teams find a majority of traffic never needed the pricier model.