Google Gemini pricing: what the API really costs

Quick answer: three different Gemini prices

People searching for Google Gemini pricing usually mean one of three different things, and Google does not make the distinction obvious. First, the consumer subscription. The Gemini app's paid plan (Google AI Pro) is a fixed fee of about $20 per month. It covers the chatbot and app features, not the API. Nothing you pay there raises your API quota by a single request. Second, the Gemini API through Google AI Studio. This is what developers actually integrate. It has a genuinely free tier with hard rate limits, and a paid tier that bills per million tokens with no monthly fee. A hobby project typically costs single-digit dollars per month; a small production app lands in the hundreds. Third, Vertex AI. Same models, enterprise wrapper, Google Cloud billing, per-token rates in the same range. If price is the question you are asking, you almost certainly want AI Studio rather than Vertex. The rest of this guide covers the API: what the free tier allows, what paid usage costs at realistic volumes, and what to do when the rate limiter starts returning 429s.

Gemini API pricing tiers and free-tier rate limits

The Gemini API meters three things at once: requests per minute (RPM), tokens per minute (TPM), and requests per day (RPD). Whichever ceiling you hit first, the API answers 429 until that window resets. Limits apply per Google Cloud project and per model, so creating a second key inside the same project changes nothing. Google revises these numbers frequently, so treat the table as the shape of the system rather than a contract. The figures are approximate Flash-class limits as of mid-2026; Pro-class models get materially lower caps at every tier. Check Google's rate-limits documentation for the exact values applied to your project today.

Free vs paid: what actually changes

Adding a billing account does more than raise the ceilings. The most overlooked difference sits in the terms: Google may use free-tier prompts and responses to improve its products, and that can include human review. Paid API traffic is excluded from product improvement. If you are prototyping on anything sensitive, that line matters more than any rate limit. The daily request cap is the other practical difference. On the free tier it is a hard ceiling that resets at midnight Pacific time. No retry strategy gets around it; you wait, upgrade, or route the traffic somewhere else.

Test your tier from the command line

One curl tells you whether your key works and which tier you are effectively on. The response includes a usageMetadata block with exact input and output token counts, which is the number you multiply by list rates when forecasting spend. If you get HTTP 429 with status RESOURCE_EXHAUSTED instead of a completion, you have either burned the day's free-tier quota or you are sending faster than the per-minute window allows.

curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.5-flash:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents":[{"parts":[{"text":"Explain token billing in one paragraph."}]}]}'

What Gemini costs per month: three scenarios

These scenarios use Gemini 3.5 Flash at its list price of $1.50 per million input tokens and $9.00 per million output tokens. Two things dominate the math. Output tokens cost six times as much as input, so verbose responses are the expensive part. And chat applications resend conversation history on every turn, so input volume grows fast with conversation length unless you trim context. If the chat-frontend row describes your workload, note that the same math applies to any model, and budget-class models such as DeepSeek V4 Flash change the totals dramatically for conversational use. Also budget separately for long-document jobs: Pro-class Gemini models bill a higher rate once a prompt crosses the large-context threshold.

Troubleshooting Gemini API rate limits

A 429 with RESOURCE_EXHAUSTED is generic, so start by working out which of the three meters you actually hit. The client below retries transient per-minute trips with backoff and gives up quickly on anything that retrying cannot fix.

• Daily cap exhausted (free tier): the RPD counter resets at midnight Pacific time. No retry pattern helps; upgrade the tier or move the traffic.
• RPM bursts: parallel workers and naive retry loops trip the per-minute limit even at low daily volume. Add exponential backoff with jitter.
• TPM trips on long context: a handful of 100K-token prompts can exhaust tokens-per-minute while your request rate looks tiny. Trim history or cache stable prefixes.
• Wrong knob: limits attach to the Google Cloud project, not the API key. Minting new keys in the same project does nothing.
• Model class: Pro-class caps are far lower than Flash-class at the same tier. Route background jobs to Flash and reserve Pro for the calls that need it.

import random
import time

from google import genai
from google.genai import errors

client = genai.Client()  # reads GEMINI_API_KEY from the environment

def generate(prompt: str, retries: int = 5) -> str:
    for attempt in range(retries):
        try:
            resp = client.models.generate_content(
                model="gemini-3.5-flash", contents=prompt
            )
            return resp.text
        except errors.APIError as e:
            if e.code != 429 or attempt == retries - 1:
                raise
            time.sleep(min(2 ** attempt + random.random(), 30))
    raise RuntimeError("unreachable")

The gateway option: fixed prices, no tier ladder

If your blocker is the tier ladder rather than the per-token price, an AI gateway is the pragmatic alternative. apisrouter.com serves gemini-3.5-flash at $1.20 per million input tokens and $7.20 per million output, with global models at 20% below official list pricing and pay-as-you-go billing instead of a subscription. There is no signup form: pay at /topup, the API key arrives by email, and your first top-up adds +100% balance. Quota becomes a balance you spend down rather than a per-minute negotiation with a tier system. The endpoint is OpenAI-compatible, so anything that accepts a custom base URL works: point it at https://api.apisrouter.com/v1 and the model list populates from /v1/models. The same catalog carries claude-sonnet-4-6, gpt-5.5, deepseek-v4-flash, and glm-5, which makes A/B-testing Gemini against alternatives a one-line change.

curl https://api.apisrouter.com/v1/chat/completions \
  -H "Authorization: Bearer $KEAIAPI_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-3.5-flash",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

FAQ