Google Gemini API pricing September 2026 — official list & per-token rates 2026

This is Google Gemini API pricing as of September 2026: Google's official per-token list price for each current Gemini model, side by side with what the same model costs on Kunavo, plus worked cost examples you can sanity-check and the cheapest way to call Gemini in production. Gemini is among the best value in frontier AI, and Kunavo prices it about 70% below Google's list rate behind one OpenAI-compatible API.

Rates last verified September 3, 2026. Kunavo's per-token prices below are read live from the model catalog, and the “Google list” column tracks Google's published rate — the official source is ai.google.dev/gemini-api/docs/pricing.

Official Google price list vs Kunavo — September 2026

Rates are per 1M tokens, in USD, as billed on Kunavo. The “Google list” column is Google's official published rate for the same model, taken from ai.google.dev on the verification date above — so this table is the official price list and the discounted one in a single view.

*Gemini 3.6 Flash (released July 2026) is the current generation and needs an honest asterisk: Google is running an introductory rate of $0.75 / $3.75 through December 31, 2026, before it reverts to the $1.50 / $7.50 list price on January 1, 2027. During the intro window Google direct is cheaper than Kunavo's $1.05 / $5.25 for this one model — if that raw rate is all you need, take it. Kunavo's case is the single OpenAI-compatible key, Stripe billing with no Google Cloud account, and one balance that also covers Claude, GPT and the image/video models; from 2027 the price is ~30% under Google's list again.

Flash is the high-volume workhorse; Pro is for harder reasoning, vision and long-context jobs. Live rates always show on the pricing page and each model page (gemini-2-5-flash, gemini-2-5-pro, gemini-3-6-flash).

How Gemini API pricing works per token

You pay for input tokens (everything you send — system prompt, retrieved context, the user message) and output tokens (what the model generates). Output is the more expensive side, so the single biggest lever on a Gemini bill is how much text you let the model write. Images and audio are converted to token-equivalents and billed on the same meter.

Worked cost examples

Real numbers at Kunavo's Gemini 2.5 Flash rate, except the last row which uses Gemini 2.5 Pro:

At those rates a 100,000-document classification batch on Flash runs about $6, and a million chatbot turns about $315. The math, runnable:

# Kunavo Gemini 2.5 Flash rates (USD per 1M tokens)
IN_RATE, OUT_RATE = 0.09, 0.75

def cost(in_tokens: int, out_tokens: int) -> float:
    return in_tokens / 1_000_000 * IN_RATE + out_tokens / 1_000_000 * OUT_RATE

print(cost(1_000, 300))            # one chatbot turn   -> $0.000315
print(cost(8_000, 500))            # one RAG answer     -> $0.001095
print(cost(500, 20) * 100_000)     # 100k-doc batch     -> ~$6.00

Kunavo pricing and Stripe billing

There is no subscription and no Google Cloud project. You top up a balance (Stripe or local payment methods), and calls draw down from it at the per-token rates above. Pay-as-you-go from a $10 minimum top-up, the balance never expires, and larger top-ups carry bonus credit. One balance covers Gemini and every other model — Claude, GPT, image, video and audio — so you are not reconciling a separate invoice per provider.

Which Gemini model should I choose?

• gemini-2-5-flash — default for chat, extraction, classification, summarization and most RAG. Fast and the cheapest capable option.
• gemini-2-5-pro — reach for it when Flash is not accurate enough: multi-step reasoning, code, vision and very long context.

A good pattern is to route by difficulty: Flash for the common case, escalate to Pro only when a check fails. See the AI cost optimization guide for the routing pattern in code, and the Claude and GPT pricing guides to compare providers.

Cutting your Gemini bill

1. Tier down. Send the easy 80% to Flash; reserve Pro for the hard 20%.
2. Cap output. Set max_tokens and stop sequences — output is the pricey side of the meter.
3. Trim input. Retrieve fewer, better RAG chunks instead of stuffing the whole knowledge base into context.
4. Batch. Group independent calls to keep latency down and avoid retry storms.

FAQ