The price list in one table
People who search for nano banana api pricing are often comparing image-generation costs, which are usually quoted per picture. This site is different. It sells access to a text language model through an OpenAI-compatible endpoint, and text is billed by the token. If you need images, this is not the right product, but it fits well next to an image tool for writing prompts, captions and scripts.
| Item | Price or rule |
|---|---|
| Input tokens | $0.15 per 1M |
| Output tokens | $0.50 per 1M |
| Subscription | None, pay as you go |
| Minimum top-up | $20 (whole amounts up to $500) |
| Failed or refused requests | Not charged |
| Credit expiry | Never |
| Trial | $0.50, one click, 7 days, no card |
The current figures live on the pricing page, and you get a key from the API key page.
What a token is, in money terms
A token is a chunk of text, often a short word or part of a longer one. As a rough rule of thumb for English, 1,000 tokens is a bit under 800 words, though this varies with the text. You are billed for two counts on every call: the tokens you send (the prompt, including any system message and conversation history) and the tokens the model writes back.
Output costs more than input, about 3.3 times as much per token ($0.50 divided by $0.15). That has a practical consequence. A request that sends a long document and asks for a one-line answer is cheap. A request that sends one sentence and asks for a long story costs more per word of text than the reverse. When you plan a budget, estimate the output length first.
The formula is simple:
cost = (input_tokens x 0.15 + output_tokens x 0.50) / 1,000,000
Every response includes a usage object with the exact counts, and the X-Balance-USD response header reports your remaining credit, so you can check the math against real calls.
Worked examples
These numbers are computed straight from the two prices.
A chat message
2,000 input tokens and 500 output tokens. Input: 2,000 x $0.15 / 1M = $0.0003. Output: 500 x $0.50 / 1M = $0.00025. Total: $0.00055. The $0.50 trial credit covers about 909 such messages, since 0.50 / 0.00055 is 909.
A prompt-writing batch for an image workflow
500 requests, each with 400 input tokens and 600 output tokens. Input: 200,000 tokens x $0.15 / 1M = $0.03. Output: 300,000 tokens x $0.50 / 1M = $0.15. Total: $0.18.
A long story with a big context
One request with 6,000 tokens of notes and a 1,500-token reply. Input: $0.0009. Output: $0.00075. Total: $0.00165. A thousand of these cost $1.65.
The worst case for one call
The context window is 100,000 tokens, prompt and completion together. Fill 96,000 with input and let the model write 4,000: input is 96,000 x $0.15 / 1M = $0.0144 and output is 4,000 x $0.50 / 1M = $0.002, for $0.0164. Even a maxed-out request costs under two cents.
An agent loop
An agent makes 50 calls, each re-sending 20,000 tokens of history and writing 1,000 tokens. Input: 1,000,000 tokens = $0.15. Output: 50,000 tokens = $0.025. Total: $0.175. This is the case that surprises people, because the history is billed again on every call. Trim old messages or summarize them to keep the input count down.
Top-ups, bonuses and how to pay
Credit is prepaid. You choose any whole dollar amount from $20 to $500. Larger top-ups add a bonus on top:
| Top-up | Bonus | Credit you receive |
|---|---|---|
| $20 | none | $20.00 |
| $50 | +5% | $52.50 |
| $100 | +10% | $110.00 |
| $200 | +15% | $230.00 |
| $500 | +15% | $575.00 |
Payment is by crypto (USDT on TRC20, USDC on Base). Since credit never expires, a bigger top-up is simply a better rate, with no deadline pushing you to spend it. A $20 top-up at the chat-message size above buys roughly 36,000 messages.
Trial credit and rate limits
To try the service first, sign up with Google or with an email and password, then press the claim button on the key page. You receive $0.50 of trial credit, one per person, valid for 7 days, no card required. A trial key allows 2 concurrent requests and 60 requests per minute.
A full key, meaning one with paid credit, allows 8 concurrent requests and 300 requests per minute. These are limits on speed, not on spend. If you exceed them you get HTTP 429 with the code rate_limited or concurrency, and a temporary overload returns 429 with Retry-After: 5. Because failed requests are not charged, a 429 never costs you money; it only means you should wait.
Use the trial for something that resembles your real traffic, not a hello-world prompt, so the usage numbers you measure tell you something about the bill.
Estimating your own bill
You can get a decent estimate in a few minutes with a short script that runs a sample of your real prompts and sums the usage.
import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.nanobananaapis.com/v1",
api_key=os.environ["NANOBANANA_KEY"],
)
IN_PRICE, OUT_PRICE = 0.15, 0.50 # dollars per 1M tokens
prompts = ["Write an image prompt for a desert market at noon.", "Write a caption for a paper boat photo."]
total = 0.0
for p in prompts:
r = client.chat.completions.create(
model="uncensored",
messages=[{"role": "user", "content": p}],
max_tokens=300,
)
u = r.usage
cost = (u.prompt_tokens * IN_PRICE + u.completion_tokens * OUT_PRICE) / 1_000_000
total += cost
print(u.prompt_tokens, u.completion_tokens, round(cost, 6))
print("sample total:", round(total, 6))
Multiply the average by your expected monthly call count, then add a margin of 20 to 30 percent for retries and longer-than-expected outputs. Set max_tokens if you want a hard ceiling on output length, though it is optional. For the setup details behind this snippet, see the docs page.
Finally, two things that are not on the bill: there is no charge for failed or refused calls, and prompts are not used for training. For a text step in an image pipeline, such as the one described in the prompt generator article, the cost per generated prompt is usually a small fraction of a cent.
Ways to keep the bill small
Because the unit is the token, most savings come from sending and receiving fewer of them. These changes help in roughly this order.
- Trim the history. In a chat or an agent, old turns are billed again with every new call. Keep the last handful of turns, and replace the rest with a short summary.
- Shorten the system message. A 1,500-token set of instructions sent 100,000 times is 150M input tokens, which is $22.50 at $0.15 per 1M. If half of it is filler, you are paying for filler.
- Ask for the length you need. Output is the expensive side. Say "two sentences" or "a JSON object with three fields" instead of letting the model ramble, and set
max_tokensas a ceiling. - Batch related questions. One request that asks for five short items usually costs less than five separate requests that each repeat the same context.
- Do not retry blindly. A 429 is free but a retry loop with no delay just burns time. Wait five seconds and try again.
- Test on a small sample. Run 20 real requests, read the usage numbers, and only then run the full set.
None of this needs special tooling. The usage object in each response is enough to find which prompt is the heavy one.
Questions to ask before you compare prices
Comparing per-token prices between services is only fair when the products are alike. Before you set two numbers side by side, check these points, and check each provider's current pricing yourself rather than trusting a table from someone else.
- Same kind of output? Text tokens and per-image prices are not comparable units. This service is text only.
- Same refusal behavior? A refused or filtered answer that you still pay for is more expensive than its list price suggests. Here, refused and failed requests are not charged.
- Prepaid or subscription? A monthly fee is a fixed cost whether you use the service or not. Prepaid credit that never expires has no such floor.
- What are the limits? A cheap token price with tight concurrency may cost you in waiting time. Here it is 8 concurrent requests and 300 per minute on a full key.
- Is the context large enough? A 100,000-token window covers long documents and long chats in one call.
When your numbers look right, create a key, spend the trial credit on your real prompts, and compare the measured cost per task with your budget.
