# What is Gemini 2.0, and what does Google mean by the "agentic era"?

Date: 2024-12-12 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-gemini-2-0-flash/

**Answer:** Gemini 2.0 is Google's new generation of AI models, introduced on December 11, 2024, starting with an experimental Gemini 2.0 Flash that is faster than Gemini 1.5 Pro, can create images and speech as well as text, and can use tools like Google Search on its own. Google calls it a model for the "agentic era," meaning AI that plans several steps and takes actions for you, under your supervision.

Yesterday, December 11, 2024, Google introduced **Gemini 2.0**, the second generation of the AI model family it launched a year ago. The first member to arrive is **Gemini 2.0 Flash**, released as an experimental version for developers and for anyone using the Gemini assistant on the web. Alongside it, Google showed several research prototypes of AI agents built on the new model.

Sundar Pichai, CEO of Google and Alphabet, framed the change simply: the first generation helped make sense of information, and the second is meant to make it more useful. The word Google keeps returning to is **agentic**.

## What Google released

"Flash" is Google's name for its fast, efficient model size, the one developers use most because it answers quickly. Google says 2.0 Flash beats the larger Gemini 1.5 Pro on key benchmarks while running at twice its speed, and TechCrunch reports that Google now treats it as the new flagship of the family.

| What's new | What it means |
|---|---|
| Native image output | It can produce pictures mixed with text in the same answer |
| Native speech output | It can read text aloud in one of eight voices, faster, slower or in a requested style |
| Native tool use | It can decide to run a Google Search, execute code or call a developer's own functions |
| Multimodal Live API | Developers can build apps that stream live audio and video to the model and talk back in real time |
| Gemini app access | A chat version of 2.0 Flash experimental appears in the model menu on desktop and mobile web |

Some parts are limited for now. Developers can send images, video and audio and get text back, but image and speech output go only to early-access partners until a wider release planned for January, when Google also promises the stable version and more model sizes. Paying Gemini Advanced subscribers also get a new feature, **Deep Research**, which explores a topic and compiles a report.

## What "agentic" means

A regular chatbot answers one question at a time. An **agent** is an AI system that takes a goal, breaks it into steps and carries them out, often by using tools. Google defines its agentic models as ones that grasp more of your surroundings, think several steps ahead and act on your behalf, with you supervising.

Here's a simple way to see the difference with a single request, "Plan a Saturday in Ottawa for two kids who love science":

| Step | A chatbot | An agent |
|---|---|---|
| 1 | Writes a plan from what it learned in training | Searches for current museum hours and events |
| 2 | Leaves checking to you | Compares travel times between places |
| 3 | Stops | Drafts the schedule, then asks you before booking anything |

Gemini 2.0 Flash is designed for the agent column: it can plan, follow complex instructions, call several tools in a row and reason across images, sound and text.

## The prototypes

Google showed four experiments built on 2.0, mostly limited to trusted testers for now:

- **Project Astra**, a universal assistant for your phone, now speaks several languages (even mixed in one sentence), uses Google Search, Lens and Maps, and remembers up to 10 minutes of the current session. A small group will test it on prototype glasses.
- **Project Mariner** works inside the Chrome browser. It reads what's on the screen, including text, images and forms, and completes tasks for you through an experimental extension. On WebVoyager, a test of real web tasks, Google reports a score of 83.5%. It acts only in the tab you have open and asks for confirmation before sensitive steps such as a purchase.
- **Jules** is a coding agent that works inside GitHub, the platform where programmers store code. It can take a reported problem, make a plan and carry it out under a developer's supervision.
- **Game agents** watch the screen while you play titles like Clash of Clans and suggest what to do next.

## How Google is handling safety

Agents that act raise new risks, and Google names one directly: **prompt injection**, hidden instructions in an email, document or web page that try to hijack the AI. It says Mariner is being trained to put your instructions ahead of such attempts. All images and audio from 2.0 Flash carry **SynthID**, Google's invisible watermark, so supported Google tools can flag them as AI-made. TechCrunch notes that Google shared no image or audio samples at launch, so the quality of those outputs can't yet be judged.

## Why it matters

This is Google's clearest statement yet that the next contest in AI is about doing, not only answering. In the past two months Anthropic has opened a beta in which Claude operates a computer, and Google now shows a browser agent and a coding agent of its own. Many expect these assistants to move from demos into products over the coming year.

For everyday users, the immediate change is smaller but real: a faster model in the Gemini app, with Google planning to bring 2.0 to more of its products early next year, starting with Search, where AI Overviews already reach a billion people.

> **Try it:** In the Gemini app on the web, choose 2.0 Flash Experimental from the model menu. Ask: "I have flour, eggs, milk and apples. Suggest three breakfasts, then tell me which needs the fewest steps." Then send the same request to the model you used before and compare speed and clarity.

## How can you try Gemini 2.0 Flash today?

If you use Gemini, open gemini.google.com on a computer or your phone's browser and pick 2.0 Flash Experimental from the drop-down menu at the top. Google says it is coming to the Gemini mobile app soon. Because it's labeled experimental, expect occasional rough edges.

If you build software, sign in to Google AI Studio, Google's website for developers to test its models, and select the 2.0 Flash experimental model. You can try tool use there, such as letting the model ground its answer in Google Search, and experiment with live voice and video through the Multimodal Live API. Companies on Google Cloud can use it through Vertex AI.

Start with tasks where you can check the outcome easily: a short research question with sources, a photo to describe, a small calculation done with code execution. That will show you quickly where the model is strong and where it still needs your eye.

## Sources

- Introducing Gemini 2.0: our new AI model for the agentic era: https://blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/
- Gemini 2.0, Google's newest flagship AI, can generate text, images, and speech: https://techcrunch.com/2024/12/11/gemini-2-0-googles-newest-flagship-ai-can-generate-text-images-and-speech/