# What is Gemini 3.5 Flash, and why is Google building it around agents?

Date: 2026-05-20 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-gemini-3-5-flash/

**Answer:** Gemini 3.5 Flash is Google's newest AI model, launched on May 19, 2026 at Google I/O as the first of the Gemini 3.5 family, and it is now the default model in Google Search's AI Mode and in the Gemini app for everyone. Google says it beats its previous top model, Gemini 3.1 Pro, on coding and agent tests while running about four times faster than other leading models, which makes it suited to AI agents that work through long, multi-step tasks.

Yesterday, Tuesday, May 19, 2026, Google opened its annual I/O developer conference by introducing **Gemini 3.5**, its new family of AI models, and releasing the first member right away: **Gemini 3.5 Flash**. It is available the same day to the public, to developers and to businesses, and it becomes the standard model, worldwide, for both the Gemini app and AI Mode, the conversational answer mode inside Google Search.

Google's own title for the family puts the emphasis on action. TechCrunch reads the launch as a shift in how Google pitches AI: less as something you chat with, more as something that plans, builds and revises real work with little supervision.

## What Google released

Google usually ships its models in sizes. **Pro** models are the largest and most capable. **Flash** models are lighter, faster and cheaper to run, and they normally trail the Pro version of the same generation. This time Google says the smaller model beats the previous generation's best.

| | Gemini 3.5 Flash | Gemini 3.1 Pro (previous top model) |
|---|---|---|
| Terminal-Bench 2.1 (multi-step command-line tasks) | 76.2% | lower, per Google |
| MCP Atlas (using outside tools through MCP) | 83.6% | lower, per Google |
| GDPval-AA (realistic office tasks, scored like a chess rating) | 1656 Elo | lower, per Google |
| Speed | About 4 times faster than other frontier models in output tokens per second | not stated |

These are Google's figures. **MCP**, the Model Context Protocol, is a common standard for connecting an AI model to outside tools and data. **Elo** is a rating system borrowed from chess, where a higher number means the model wins more head-to-head comparisons. A **token** is a small piece of text, often part of a word, so "tokens per second" measures how fast the model writes.

Koray Kavukcuoglu, Google DeepMind's chief technologist, told reporters that the model outperforms 3.1 Pro on nearly all benchmarks, and that Google also built an optimized version that runs 12 times faster with the same quality. Gemini 3.5 Pro is already in use inside Google, and the company plans to release it next month.

Where you can find 3.5 Flash today:

- **Everyone:** the Gemini app, plus AI Mode in Search, where it is now the default.
- **Developers:** the Gemini API in Google AI Studio, Android Studio, and Antigravity, Google's coding environment built around agents.
- **Businesses:** Gemini Enterprise and the Gemini Enterprise Agent Platform.

## Why speed matters for agents

An **AI agent** is a model that works toward a goal over many steps, using tools such as a code editor, a browser or a spreadsheet, rather than replying once. Agent work multiplies the number of times a model is called. If one task needs hundreds of steps, a model that answers four times faster finishes far sooner and usually costs less. Google says 3.5 Flash often completes long tasks at under half the cost of other frontier models.

Google also leans on **subagents**: several copies of the model, each given one part of a big job, working at the same time and then combining their results. Think of a project lead handing chapters of a report to different writers. Google's planned pairing makes this explicit. Tulsee Doshi, a senior director and head of product at Google, told TechCrunch that 3.5 Pro will act as the planner while Flash handles the many sub-tasks.

The model was developed alongside Antigravity, which Google updated at I/O as a stand-alone desktop app, version 2.0. The surrounding software, sometimes called a **harness**, gives agents a workspace where they can run code, check results and coordinate. In one demonstration Google published, two agents read a famous research paper about the game-playing system AlphaZero and built a playable game in six hours. Onstage, Google engineer Varun Mohan showed agents splitting up the work of building an operating system.

Doshi added that the model can work on its own for hours but will stop and ask the user when it reaches a decision or permission question that needs human judgment.

## A worked example

Here is the kind of request that suits a fast agent model, written as steps rather than a single question:

```prompt
I run a small bakery. Attached is a spreadsheet of last year's daily sales.
1. Clean the data: remove duplicate rows and flag missing days.
2. Find our three slowest months and our best-selling items in each.
3. Suggest two promotions for the slowest month, with a reason for each.
4. Make a one-page summary with a simple chart.
Ask me before making any assumptions about prices.
```

Each step depends on the one before, and the model has to keep track of the data throughout. That chain of dependent steps is exactly what Google is optimizing for.

> **Try it:** Open the Gemini app, give it a small multi-step job like the bakery example using your own household budget or a club's event list, and watch whether it completes every step in order and asks before guessing.

## Why it matters

For everyday users, the change is quiet but wide: 3.5 Flash now answers in the Gemini app and in Search's AI Mode, so many people are using it without choosing anything. It also powers **Gemini Spark**, a new personal agent meant to run around the clock and act on your behalf under your direction. Spark is starting with trusted testers, with a beta planned next week for Google AI Ultra subscribers in the US.

For businesses, Google names early partners. Shopify runs parallel subagents to forecast merchant growth, Macquarie Bank is testing faster customer onboarding from documents of more than 100 pages, and Salesforce is building it into its Agentforce product.

On safety, Google says it developed the model under its Frontier Safety Framework, its internal rules for serious risks, with stronger protections in cybersecurity and in chemical, biological, radiological and nuclear topics. It also says the model refuses safe questions by mistake less often, and that it now uses interpretability tools, methods for inspecting a model's internal reasoning, to check responses before they go out.

The launch arrives less than a month after OpenAI's GPT-5.5, which was also pitched around agents. The leading labs now agree on the direction; the contest is over speed, cost and reliability.

## How can a beginner start using Gemini 3.5 Flash today?

The simplest way is the Gemini app or website: it now runs on 3.5 Flash by default, so you need no setting changes. AI Mode in Google Search uses it too, which makes it handy for questions that need several steps of research.

To get the most out of it, write requests as a short plan: the goal, the material you are giving it, the steps and the point where it should stop and ask you. Check its output, especially numbers and dates, before you rely on it.

If you are curious about building, Google AI Studio lets you try the model through the Gemini API in your browser, and Antigravity is aimed at developers who want to experiment with coding agents.

## Sources

- Gemini 3.5: frontier intelligence with action: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/
- With Gemini 3.5 Flash, Google bets its next AI wave on agents, not chatbots: https://techcrunch.com/2026/05/19/with-gemini-3-5-flash-google-bets-its-next-ai-wave-on-agents-not-chatbots/