Gemini API Pricing 2026 — Free Tier, Model Prices, and Gemini 3.8 Flash

Gemini API costs are billed by token, with no monthly fee. To try it for free, Google AI Studio has a rate-limited free tier; for production, you pay per 1M tokens, with separate rates for input and output. As of October 5, 2026, Kunavo’s Gemini 3.8 Flash costs $0.525 for input / $2.625 for output — Google currently charges a limited-time promotional price of $0.75 / $3.75 for its 3.x Flash models (through December 31, 2026; from 2027, the standard price returns to $1.50 / $7.50), and Kunavo is about 30% below that promotional price. The previous-generation Gemini 2.5 Flash costs $0.09 / $0.75, and Gemini 3.1 Pro costs $0.70 / $4.20.

There’s one exception to mention first: Gemini 3.6 Flash. On Kunavo it costs $1.05 / $5.25, approximately 40% higher than Google's discounted price — during the promotional period, calling Google directly is cheaper. 3.7 and 3.8 Flash are both newer and cheaper than it, So don’t choose 3.6 based on price.

Rates verified on October 5, 2026. See ai.google.dev/gemini-api/docs/pricing for Google’s official prices and free-tier details; Kunavo’s rates are shown in our model catalog and match the amounts billed by the API.

“Gemini cost” actually means three different things

The most common misconception is in the middle row: a Google AI Pro subscription does not give you more API credits. The app’s monthly fee and the developer API are billed separately. To make calls from code, you still need to use the free tier or pay by token. Conversely, if you only chat in the Gemini app and don’t write code, the free version or a subscription is enough; you don’t need the API.

Is the Gemini API free? What can you do with the free tier?

There is a free tier, and it really costs $0: create a key in Google AI Studio and calls within the rate limits are free. It’s more than enough for trying models and building prototypes. Google adjusts the specific limits, so this page doesn’t repeat figures that may become outdated — see Google’s pricing page and rate limits documentation.

Its limits are also clear: it can’t handle production traffic, and the figures can change at any time, so you need to pay for production. Kunavo has no free tier, but also has no monthly fee — the minimum prepaid balance is $10, you pay for what you use, and months with no calls cost $0.

Gemini API prices — per 1M tokens by model

Those three rows for 3.x Flash compare with the promotional prices Google is actually charging today, not its standard prices for 2027. Compared with the standard prices, the figures would look much better, but those aren’t what you’re paying Google now. Gemini 3.1 Pro has another pricing rule: when a single request’s prompt exceeds 200K tokens, the entire request is billed at 2× the input rate / 1.5× the output rate. Specifications for each model are on the Gemini 3.8 Flash, Gemini 3.1 Pro, and Gemini 2.5 Flash model pages; all rates are on the pricing page.

Gemini 3.8 Flash pricing — what one request actually costs

A rate table is less useful than seeing the cost of one request. The examples below are all calculated from the rates above:

One thing that can push up your bill: thinking tokens. Gemini 3.x Flash and 3.1 Pro both “think” before answering. Thinking tokens don’t appear in the response, but are billed at the output rate. They aren’t included in the table above, so difficult tasks will cost more than the table shows — use usage for the number of output tokens in the response. Calculate using your own numbers:

# Gemini API 的費用，直接用實際 token 數算。
# 費率為每 1M token USD：(input, output)
RATES = {
    "gemini-3-8-flash": (0.525, 2.625),
    "gemini-3-1-pro":   (0.70, 4.20),
    "gemini-2-5-flash": (0.09, 0.75),
}

def cost(model, inp, out):
    i, o = RATES[model]
    return inp / 1e6 * i + out / 1e6 * o  # 思考 token 按輸出費率計

print(cost("gemini-3-8-flash", 8_000, 500))         # RAG 回答一次
print(cost("gemini-2-5-flash", 500, 20) * 100_000)  # 10 萬筆分類

Call Gemini with your existing OpenAI SDK

Kunavo is an OpenAI-compatible endpoint: replace base_url and the key, set model to the Gemini model, and you’re ready to go. You don’t need a Google Cloud project or billing account. The same key can also call Claude, GPT, and image and video models.

from openai import OpenAI

client = OpenAI(
    api_key="sk-kn-...",
    base_url="https://api.kunavo.com/v1",  # 只改這一行
)

resp = client.chat.completions.create(
    model="gemini-3-8-flash",
    messages=[{"role": "user", "content": "用三句話說明什麼是 token"}],
)
print(resp.choices[0].message.content)
print(resp.usage)  # 這次請求計費用的 token 數

For the full steps, see Quick start. To get a key directly from Google, see How to get a Gemini API key (English).

Honestly: when it’s better to go directly to Google

• You need guaranteed quota or an SLA. The gateway uses shared capacity: there is no dedicated quota or contractually guaranteed SLA. If you’re at a scale that needs either, apply directly to Google.
• You can only use Gemini 3.6 Flash. During the promotional period, Google charges less than Kunavo.
• You only chat in the app. If you don’t write code, the free version or a Google AI subscription is enough; you don’t need the API.

This route is for using one key to call Gemini, Claude, and GPT, avoiding Google Cloud billing, or keeping costs in check with a prepaid balance.

Paying from Taiwan

Paying Google directly requires enabling billing for the project in Google Cloud and linking an international credit card. Kunavo balance uses Stripe and supports cards including JCB (Visa, Mastercard, American Express, JCB, UnionPay), Apple Pay, and Google Pay; it is prepaid, with a minimum top-up of $10 and no balance expiration, and failed requests are not charged. Top-ups of $100 or more receive an additional bonus ($100 credited as $110).

Taiwan has no local payment options — JKoPay and LINE Pay are not on the supported list, so an international card is the only option. See the billing documentation for details.

How to reduce Gemini costs

1. First, choose the right model. If you want a 3.x generation model, use 3.8 or 3.7 Flash: they’re newer and cheaper than 3.6; Assign large volumes of simple tasks such as classification and extraction to 2.5 Flash.
2. Ask the model to write less. Output costs more than input, so specifying the output format and length in your prompt has the most direct effect.
3. Include less context. Retrieve a few relevant RAG chunks instead of sending the entire knowledge base — the full input is billed every time.
4. Keep 3.1 Pro prompts at or below 200K tokens. Above that, the entire request is billed at 2× the input rate / 1.5× the output rate.

For routing patterns, see the AI cost optimization guide (English). To compare other providers using the same method, see OpenAI API pricing and Claude pricing.

Frequently asked questions