How do you try Gemini 2.0 Flash in Google AI Studio?

Google released Gemini 2.0 Flash as an experimental model on December 11, and the quickest place to try it is Google AI Studio, Google's browser workspace for people who build with Gemini. You don't need to install anything or write code to start. AI Studio is a playground: a chat window on one side, settings on the other, and a button that turns whatever you just did into code you can reuse.

"Experimental" matters here. Google calls this a preview: it can change, some features are held back, and a stable version is planned for early next year. That makes it a good moment to explore what the model can do, and a poor moment to build anything you depend on.

Getting set up

1. Go to aistudio.google.com and sign in with a Google account.
2. Open a new chat prompt.
3. In the model menu on the right, choose Gemini 2.0 Flash Experimental. In code, the same model is called gemini-2.0-flash-exp.
4. Leave temperature (the setting that controls how varied the answers are) at its default for now.

A note on what you won't find yet. Google says 2.0 Flash can produce images and spoken audio of its own, but those two outputs are limited to a small group of early testers. Everyone else gets text out, with text, images, video and audio accepted as input. If you only want a chat assistant, the Gemini app also offers a chat-tuned version of 2.0 Flash Experimental in its model menu.

A worked session in three steps

Step 1: an image question with boxes. Take a phone photo of a cluttered desk or kitchen counter, upload it, and send:

List every object you can identify in this photo. Then draw bounding boxes around the three smallest objects and label each one.

Bounding boxes are rectangles that mark where an object sits in an image. Google says 2.0 Flash is better at spotting small items in busy scenes than its earlier models, and AI Studio draws the boxes right over your picture, so you can see at a glance whether it found the paper clip or mistook it for a key.

Step 2: a question that needs today's information. In the tools section of the settings panel, turn on grounding with Google Search, then ask:

What did Google announce about Gemini 2.0 this month? Answer in five bullet points and show the sources you used.

With Search available as a tool, the model decides on its own when a web lookup is needed, and the answer comes with links you can open. Turn the tool off and ask again: the difference shows you what the model knows from training and what it fetched.

Step 3: take the code with you. Click Get code. AI Studio writes out the request you just made. Google ships a new Python library alongside the model, installed with pip install google-genai, and a minimal call looks like this:

from google import genai

client = genai.Client(api_key="YOUR_API_KEY")   # create a key in AI Studio
response = client.models.generate_content(
    model="gemini-2.0-flash-exp",
    contents="Explain bounding boxes to a 12-year-old in three sentences.",
)
print(response.text)

Keep the key out of shared files; an environment variable is the safer place for it.

What's new compared with earlier Gemini models

Google's claim is speed with better quality: 2.0 Flash is described as twice as fast as Gemini 1.5 Pro while scoring higher on its key tests. The other big change is native tool use. The model was trained to call Google Search, run code and use functions that developers define, and it can combine several of them in one request. Google frames this as groundwork for agents, programs that take steps on your behalf, and it says those agents should work under human supervision.

For a beginner, the practical takeaway is simpler: one model now handles a photo, a document, a web question and a coding request in the same chat, so you can explore all of them in a single afternoon.

What can it do with live video and screen sharing?

The most striking part of the release is the Multimodal Live API, a new way to talk with the model in real time. Instead of sending a message and waiting, you stream your voice, your camera or your screen to it, and it answers out loud as you go. You can interrupt it mid-sentence, it notices when you start and stop speaking, and it remembers what it saw and heard earlier in the same session. AI Studio has a live streaming area where you can try this without code.

A good first test is screen sharing. Open a spreadsheet or a settings page you find confusing, share that window, and ask the model to walk you through it step by step. Or point your camera at a houseplant and ask what it might need. Because it keeps the session in memory, you can say "go back to the second thing you mentioned" and it will.

Keep the limits in mind. At launch the live mode works in English only, and Google notes that audio input or output makes the model less reliable at calling functions. It's an experimental preview, so expect rough edges. And treat sharing like sharing with a stranger: close email, banking tabs and anything private before you start, because the model sees whatever is on the screen you share.