On this page
- Which Free Models Actually Work (Tested September 29, 2026)
- We benchmarked it: fast to generate, very slow to start
- Get Your Free API Key
- Using the API with Python
- What Gemini Does Beyond Text
- Connect Gemini to OpenClaw (Free AI Agent)
- Gemini vs Other Free AI APIs
- Important Things to Know
- Frequently Asked Questions
- Final Thoughts
gemini-3.5-flash or gemini-3.5-flash-lite — on a new key created in September 2026, every gemini-2.5-* ID returns 404 "no longer available to new users", and the Pro models return 429 because the free tier gives them no quota. Get a key at aistudio.google.com. Pick Gemini for long documents and multimodal input; pick Groq when you need an answer to start in under a second.Most free AI APIs handle text only. Gemini adds a million-token window, built-in Google Search grounding, and multimodal input on the free tier. This guide covers which model IDs actually answer on a free key today, getting your key, first code, and how it compares.
Which Free Models Actually Work (Tested September 29, 2026)
We called every Gemini text model with one short generateContent request from a free key created in September 2026. The model list itself is misleading: GET /v1beta/models still returns the 2.5 series, but calling them fails.
| Model ID | Result on a new free key | Use it for |
|---|---|---|
| gemini-3.5-flash | Works (occasional 503 "high demand" — retry) | General use, coding, reasoning |
| gemini-3.5-flash-lite | Works, answered every call | High-volume tasks, quickest replies |
| gemini-3.1-flash-lite | Works | Lightweight tasks |
| gemini-3-flash-preview | Works | Preview model — may change or be removed |
| gemini-3.8-flash | Works, but slow (~30 s for a one-word reply) | Newest Flash; batch work only |
| gemini-flash-latest / gemini-flash-lite-latest | Work | Aliases that move to newer models without warning — pin a version in production |
| gemini-3.7-flash | Answered once, then 503 "high demand" on repeat tests | Not reliable on the free tier right now |
| gemini-pro-latest, gemini-3.1-pro-preview | 429 on the first call (free-tier input-token quota) | Paid tier only in practice |
| gemini-2.5-pro / 2.5-flash / 2.5-flash-lite | 404 "no longer available to new users" | Nothing — replace them in old code |
All the working text models report a 1,048,576-token input limit and 65,536 output tokens. The 1M window fits roughly 700,000 words — about 1,500 pages — in a single request.
Rate limits: Google no longer publishes free-tier request numbers. Its rate-limits page says limits "can be viewed in Google AI Studio," and they are set per project. Any RPM/RPD table you find elsewhere is either out of date or describes someone else's project. Check yours in AI Studio on your project's usage page. Daily limits reset at midnight Pacific Time.
If your old code breaks: a 404 that says "no longer available to new users" is not a bad key. Swap gemini-2.5-flash for gemini-3.5-flash and gemini-2.5-flash-lite for gemini-3.5-flash-lite; nothing else in the request needs to change. Keys created before the cutoff may still reach the 2.5 models, which is why older tutorials still work for their authors.
We benchmarked it: fast to generate, very slow to start
We ran Gemini through our own open benchmark alongside every other free AI API, from a US machine, same prompt, same method (August 2026, with gemini-3.6-flash, the default model at the time):
| Provider | Time to first token | Generation speed |
|---|---|---|
| Groq (gpt-oss-120b) | 0.38 s | 527 tok/s |
| Mistral (small) | 0.43 s | 167 tok/s |
| Gemini 3.6 Flash | 34.5 s | 141 tok/s |
| OpenRouter (Nemotron) | 2.6 s | 38 tok/s |
| GLM 4 Flash | 2.6 s | 20 tok/s |
That 34.5-second first-token figure is not a typo, and it is the single most important thing to know before you build on Gemini's free tier.
Why the wait: "Flash" now thinks before it answers
Gemini's current models reason internally before emitting a visible token. Those thinking tokens are billed and timed like any others — they just don't appear in the response. So the model spends half a minute deliberating on a question that Groq starts answering in under half a second.
Once it starts, it's respectable: 141 tok/s is second only to Groq among the APIs we measured. The problem is entirely the wait to begin.
What this means in practice:
- Don't use it for anything interactive. Chat, voice, autocomplete — a 30-second silence before the first word is unusable. Groq starts ~90× faster.
- Do use it for batch and background work. Summarising documents, offline analysis, anything where you wait on a result anyway. The reasoning that costs you latency also buys you better answers on hard problems.
- Watch your token budget. Thinking tokens count against your quota even though you never see them, so a "short" answer can consume far more than its visible length suggests.
Cross-platform numbers for every free API we tested: free AI APIs benchmarked.
Get Your Free API Key
- Go to aistudio.google.com and sign in with Google
- Click "Get API key", then "Create API key" and name it
- Copy and save your key — no credit card needed
Using the API with Python
Official Google GenAI SDK (recommended)
pip install google-genai
from google import genai
client = genai.Client(api_key="YOUR_GEMINI_API_KEY")
response = client.models.generate_content(
model="gemini-3.5-flash",
contents="Explain how transformers work in deep learning"
)
print(response.text)
Multimodal is just as short — upload a file and pass it alongside your prompt:
image = client.files.upload(file="photo.jpg")
response = client.models.generate_content(
model="gemini-3.5-flash",
contents=[image, "Describe what you see in this image"]
)
OpenAI-compatible endpoint
Already on the OpenAI SDK? Change three things: api_key, base_url, and model.
from openai import OpenAI
client = OpenAI(
api_key="YOUR_GEMINI_API_KEY",
base_url="https://generativelanguage.googleapis.com/v1beta/openai/"
)
response = client.chat.completions.create(
model="gemini-3.5-flash",
messages=[
{"role": "system", "content": "You are a helpful coding assistant."},
{"role": "user", "content": "Write a Python function to merge two sorted lists"}
]
)
print(response.choices[0].message.content)
What Gemini Does Beyond Text
Unlike most free APIs, Gemini handles many input and output types: image understanding (photos, diagrams, charts), audio transcription (MP3, WAV, FLAC), video analysis via the Files API, PDF parsing, sandboxed code execution, Google Search grounding, structured JSON output. Free keys also list image-generation models (the "Nano Banana" family) and text-to-speech models; whether your project gets free quota for them shows in AI Studio.
Connect Gemini to OpenClaw (Free AI Agent)
Pair Gemini's free API with OpenClaw to build an agent that browses the web, manages files, and chats on WhatsApp or Telegram.
npm install -g openclaw@latest
openclaw onboard
When prompted, select Google (Gemini) and enter your key. Choose gemini-3.5-flash for everyday use or gemini-3.5-flash-lite for quicker turns (the Pro models have no free quota). To configure manually, edit ~/.openclaw/openclaw.json:
{
"models": {
"mode": "merge",
"providers": {
"gemini": {
"baseUrl": "https://generativelanguage.googleapis.com/v1beta/openai/",
"apiKey": "GEMINI_API_KEY",
"api": "openai-completions",
"models": [
{
"id": "gemini-3.5-flash",
"name": "Gemini 3.5 Flash",
"reasoning": true,
"input": ["text", "image"],
"contextWindow": 1048576,
"maxTokens": 65536
}
]
}
}
},
"agents": {
"defaults": {
"model": { "primary": "gemini/gemini-3.5-flash" },
"models": { "gemini/gemini-3.5-flash": {} }
}
}
}
Gemini vs Other Free AI APIs
| Feature | Google Gemini | Groq | DeepSeek | Alibaba Bailian |
|---|---|---|---|---|
| Top free model | gemini-3.5-flash | gpt-oss-120b | deepseek-v4-flash | Qwen 3.6-Plus |
| Context Window | 1M tokens | 128K | 128K | 1M |
| Free RPD | Not published (see AI Studio) | 1,000 (chat models) | None — top-up required | 1M tokens/model |
| Multimodal | Text+Image+Audio+Video | Text only | Text only | Text+Image |
| OpenAI Compatible | Yes | Yes | Yes | Yes |
| Credit Card | No | No | No | No |
| Speed | Slow to start (34.5 s first token), 141 tok/s after | 0.38 s first token, 527 tok/s | Moderate | Fast |
| Built-in Search | Yes (free) | No | No | No |
Important Things to Know
- Data usage: On the free tier, Google may use your prompts and responses to improve its products. Don't send sensitive data.
- Age requirement: Must be 18+ to use AI Studio.
- EU restriction: Apps for EU/EEA/UK users must use the paid tier.
- Deprecated SDK: The old
google-generativeaipackage is deprecated — usegoogle-genai.
Frequently Asked Questions
Does Gemini's free API need a credit card?
No. You get a key at AI Studio with just a Google account — no card, no billing setup. You only move to the paid tier if you exceed the free rate limits or build for EU/EEA/UK users.
Which free model should I use?
Start with gemini-3.5-flash. If it returns 503 "high demand" or you need quicker replies, use gemini-3.5-flash-lite, which answered every call in our September 29, 2026 test. Don't use gemini-2.5-* IDs: new keys get a 404. The Pro models (gemini-pro-latest, gemini-3.1-pro-preview) returned 429 on the first call because the free tier gives them no quota. Google doesn't publish free request limits, so check yours in AI Studio.
Why does gemini-2.5-flash return 404?
Google closed the 2.5 models to new users. The error reads "This model models/gemini-2.5-flash is no longer available to new users", even though the model still appears in GET /v1beta/models. Change the model ID to gemini-3.5-flash; the rest of your code works unchanged.
Can I use the OpenAI SDK with Gemini?
Yes. Point the OpenAI SDK at https://generativelanguage.googleapis.com/v1beta/openai/, pass your Gemini key, and set the model to a Gemini ID. Existing OpenAI code works with three changed values.
Is my data private on the free tier?
No. Google may use free-tier prompts and responses to improve its products, so don't send sensitive data. For confidential workloads, use the paid tier.
Final Thoughts
Choose Gemini's free tier when the job involves long documents or images, audio and video: the 1M-token window and multimodal input are its real advantages, and no credit card is required. Choose Groq for anything interactive, because Gemini's Flash models think for seconds before the first token. Whichever you pick, use a current model ID (gemini-3.5-flash or gemini-3.5-flash-lite) — many Gemini tutorials still show 2.5 IDs that fail on new keys. Get started: aistudio.google.com