Gemini 3.5 Flash is launched at Google I/O 2026 with a focus on AI agents and programming.

A Google announced, this Tuesday (19), the Gemini 3.5 Flash during the annual conference Google I/O 2026, conducted in Mountain View. The new artificial intelligence model promises superior performance to Gemini 3.1 Pro in programming and autonomous agents, in addition to operating up to four times faster than other frontier models. The company also revealed a complete redesign of the application. Gemini:, the personal agent Gemini Spark, the video creation model Gemini Omni and changes to subscription plans. Check them out below.

What is Gemini 3.5 Flash?

O Gemini 3.5 Flash It is the first model of the new family. Gemini 3.5 and it has already replaced the previous model as the application's default. Gemini: and AI mode in search of Google worldwide. In practice, anyone who uses the Gemini: Starting today, even the free version of the app will be interacting with it.

To understand what changes, a quick explanation is worthwhile. In the universe of AI models, there are different model "sizes." The largest (called Pro or flagship Typically, they tend to be smarter, but also slower and more expensive. Smaller ones (like those in the Flash line) are faster and cheaper, but may not deliver as high quality.

What Google affirms with the 3.5 Flash The difference has decreased drastically (at least until the next Pro arrives). The model delivers performance comparable to larger and more expensive models, both from the same manufacturer. Google how many competitors like OpenAI and anthropic, maintaining the speed of the Flash line.

According to the company, the 3.5 Flash It is the strongest model of Google on two specific fronts. The first is for code, that is, the ability to write, correct, and maintain programming code. The second is agentic tasks, which is when AI not only answers a question but takes a sequence of actions to solve a complete problem.

A practical example would be asking AI to analyze a financial spreadsheet, identify inconsistencies, cross-reference it with email data, and generate a final report, all without human intervention at each step.

The model features a context window of 1 million tokens (the equivalent of processing about 750 words at once, something like ten entire books), supports multimodal inputs (text, image, audio and video) and is already available globally via Gemini API, Google AI Studio, Android Studio, the agent platform Google Antigravity e Gemini Enterprise.

Performance in benchmarks

To support these claims, the Google released results in several benchmarksThese are standardized tests used by the industry to measure the performance of AI models in specific tasks. The most relevant tests of Gemini 3.5 Flash They focus on programming and automation.

The main figures presented by the company were 76,2% No. Terminal-Bench 2.1 (a test that measures the AI's ability to solve real-world tasks in a programming terminal, such as installing packages, fixing errors, and manipulating files) 83,6% No. MCP Atlas (evaluates the coordinated use of multiple tools by AI agents) and 84,2% No. CharXiv Reasoning (Test of comprehension of graphs and figures in scientific articles).

The model also reached 1.656 Elo No. GDPval-AA, a benchmark that simulates tasks of real economic value, such as the analysis of financial documents.

According to the website LLM Stats, the model surpasses the Gemini 3.1 Pro in benchmarks that resemble real-world work, such as programming, tool usage, and financial automation. However, the 3.5 Flash it is behind 3.1 Pro in benchmarks of pure academic reasoning, such as the Humanity's Last Exam (40,2% versus 44,4%) and the ARC-AGI-2 (72,1% versus 77,1%).

In practice, this means that for complex questions in science or abstract mathematics, the Pro model remains superior, but for day-to-day work tasks, such as writing code or automating processes, the 3.5 Flash takes advantage.

Furthermore, Google He emphasized that the model generates text. four times faster that other top-of-the-line models on the market

The competitive context

The launch comes at a time of pressure for the GoogleDespite having far greater financial resources than its rivals, the company has fallen behind in the race for AI applied to business. According to... Financial TimesAnalysts estimate that... Google holds between 10% and 15% of the developers and automation with AIwhile smaller competitors such anthropic e OpenAI dominate approximately 40% together.

A anthropic, creator of Claude, informed investors that it is on track to generate revenue. US$45 billion annualized, a fivefold increase compared to the end of last year, driven primarily by the popularity of its coding tools.

already the application ChatGPT, OpenAIHas 900 million weekly users, while the Gemini: It reached the same mark, but in users. monthly, which indicates lower engagement per session.

The CEO of Google, Sundar Pichai, addressed the point directly during the conference, saying that the company wants to expand the use of agents beyond the corporate and developer audience.Those who watched the presentation noticed how the Google is betting and doubling the bet in agents now.

How much does it cost for developers (and what has changed in price)?

A Google positions the Gemini 3.5 Flash as a cost-effective option for companies, stating that it performs agentic tasks. "for less than half the cost of other top-of-the-line models"But this comparison needs important context.

First, a quick explanation. Tokens These are the units of text that AI models process. Each word in Portuguese is equivalent to approximately 1,5 tokens. When a company uses the Gemini API (the interface that allows integrating AI into its own systems), it pays per million tokens processed, both in input (what you send to the model) and in output (what the model responds).

The price of Gemini 3.5 Flash in the API it is from $1,50 per million entry tokens e $9 per million exit tokens. compared to Gemini 3.1 Pro (US$ 2,50/US$ 15), the new model is in fact about 40% cheaper In both directions.

However, when we compare it to the previous Flash version, in this case the Gemini 3 Flash, which many developers currently use in production, costs $0,30 per million entry tokens e $3,00 per million exit tokens.

Therefore, for teams that were already using the 2.5 Flash In high-scale pipelines, the migration to 3.5 Flash This represents a significant increase in cost, even though the performance is considerably superior.

Córdoba	Entry (per 1M tokens)	Exit (per 1M tokens)
Gemini 2.5 Flash	US$ 0,30	US$ 2,50
Gemini 3.5 Flash	US$ 1,50	US$ 9,00
Gemini 3.1 Pro	US$ 2,50	US$ 15,00

Antigravity 2.0 and the demo that "ran Doom"

One of the highlights of Google I/O It was the presentation of Antigravity 2.0, Google's development platform designed from the ground up to create and orchestrate AI agents.

Varun Mohan, the platform's head, took to the stage and demonstrated the system's power with an ambitious proposal, as the team put the Antigravity and its agents to build the core of an operating system from scratch.

The numbers released by Google In this process, they impress. Creating the operating system used 93 subagents (smaller agents, each responsible for a specific task, such as writing a system module, testing a component, or resolving a dependency) working in parallel.

According to Google, the AI ​​agents generated 2,6 billion tokens and completed the functional core of the operating system in about 12 hours with Estimated cost less than US$1.000During the on-stage demonstration, the system initially failed to run the game Doom due to missing keyboard drivers, but the problem was fixed in real time by the agent. Antigravity, making the game playable live.

It's worth noting that, according to sources, the operating system created is not comparable to a... Windows ou Linux This is a complete system. It's merely a functional core of an experimental system, sufficient to demonstrate booting and execution.

Similarly, the "Doom" presented is more of a clone of the original code for the system, not necessarily the original executable. Still, the demo is definitely impressive.

Gemini Spark and the new look of the app.

In addition to the model, the Google It introduced significant changes to the user experience. The application Gemini: now has more than 900 million monthly active users in more than 230 countries and 70 languages, more than double the 400 million reported in I / O from last year. According to Pichai, the number of daily consultations has increased sevenfold in that period.

Neural Expressive: the new design

The app redesign adopts a visual language called Neural ExpressiveWith fluid animations, revamped typography, vibrant colors, and haptic feedback, the most noticeable change is in how Gemini presents its responses.

Until now, the experience was similar to other AI chatbots, with responses delivered as long blocks of running text. However, with the Neural ExpressiveGemini will then assemble visually tailored answers to the type of question, or at least that's what Google promises.

If the user asks about the best travel destinations, for example, the answer might come with... integrated images, interactive maps e narrated videosInstead of just a list of names, the most important information is highlighted at the top, and the rest is organized into visual layers that the user can explore as needed.

O Gemini LiveThe voice chat feature has also been integrated directly into the main chat interface. Now the user can switch between typing and voice chat without changing screens. Google also announced future support for... regional dialects in Gemini's voice. The redesign is now available globally on Android, iOS and website.

Gemini Spark: a 24/7 personal agent

O Gemini Spark This is possibly the most ambitious new product of the event for the average consumer. It's an AI personal assistant that runs 24 hours per day in dedicated virtual machines of Google Cloud, using the Gemini 3.5 Flash like an engine and the Antigravity as an execution platform. Unlike the current Gemini, which responds when asked, the Spark operates in a way proactive and continuous, performing tasks in the background under the user's direction.

The user can set up recurring tasks, such as asking... Spark all with Automatically analyze your credit card statement every month. and to flag new charges or hidden fees.

It's also possible to teach him to monitor his inbox for... updates from children's schoolExtract important deadlines and send a consolidated summary daily to the user and the partner.

In more complex flows, the Spark can cross-reference meeting notes scattered across emails and chats.to create an organized document in Google Docs ...and even draft the project follow-up email.

O Spark it initially integrates with the company's own services. Googleas the gmail, Docs and other apps from workspace, and should expand to third-party tools via MCP (Model Context Protocol) throughout the summer.

On mobile, the agent will have a dedicated interface called Android Halo, which shows live updates and the progress of ongoing tasks. A Google It reported that high-impact actions, such as sending emails or making purchases, They require user confirmation before they can be executed.

The agent will begin rolling out to trusted testers this week and should arrive as beta for subscribers of Google AI Ultra in the United States next week.

Gemini Omni and video creation

Another new feature announced alongside 3.5 Flash it was the Gemini OmniOmni is a model focused on creating and editing video directly within the Gemini app. Omni combines Gemini's intelligence with generative media models to transform... text, images and videos in high-quality visual content.

In practice, the user can send a clip recorded on their phone and edit it using voice or text commands, applying effects such as cinematic zoom, scene changes, or lighting adjustments. Google also demonstrated the creation of a... personalized AI avatar which replicates the user's appearance and voice, allowing them to be inserted into generated scenes.

O Gemini Omni It starts being released today for subscribers of the plans. AI Plus, Pro e Ultra all around the world.

New subscriptions and prices for consumers

Another noteworthy new feature is that the Google has restructured its subscription plans. The main new feature is a new plan. AI Ultra for $100 per monthThis plan is geared towards advanced developers and creators, with a usage limit five times greater than the Pro plan. The old plan AI Ultra, which cost US$ 250, fell to $ 200 per month, maintaining a limit 20 times higher than the Pro, Project Genie, YouTube Premium e 30 TB of storage.

The change puts the Google on a direct parity with competitors. A OpenAI charges US$200 for the premium plan of ChatGPT, Ea anthropic charges US$100 and US$200 for the higher price ranges of ClaudeThe difference is that the Google includes extra services in the package such as YouTube and cloud storage.

No BrazilThe new prices are already reflected on the page. Google One. The plan AI Ultra It now appears in two tracks: $ 779,90 / month (with 20 TB of storage) and $ 999,90 / month (with 30 TB), values ​​below 1.209,90 BRL Previously charged for the single plan. The other plans remain available in the country: AI Plus by $ 24,99 / month, Premium AI by $ 49,99 / month e AI Pro from $ 96,99 / month.

The company also replaced the daily prompt limits with a system called "compute-used"This measures the complexity of each request instead of counting messages. A simple question consumes little quota, while video generation or long programming sessions consume more. The limits are renewed every [time period]. five hoursAnd if the user exhausts the quota in the larger models, the system automatically migrates to smaller models instead of blocking access.

For those who don't want to pay, the Gemini 3.5 Flash It is also available for free in the Gemini app and in Google Search's AI Mode, with stricter usage limits.

Enhanced security

A Google also highlighted improvements in safety protections do Gemini 3.5The model was developed following the Frontier Safety Framework from the company, with reinforcement of mechanisms against the generation of hazardous content in the areas. cybernetics e CBRNMore (chemical, biological, radiological and nuclear).

According to the company, the new protections include tools for interpretabilityIn practice, this allows engineers to "look inside" the AI's reasoning before it provides an answer, verifying whether the model is following a safe line of thought.

The goal is to reduce both the generation of harmful content and the... incorrect refusals safe questions, a common problem in previous models, which sometimes refused to answer harmless questions out of excessive caution.

Gemini 3.5 Pro arrives next month.

A Google also confirmed that the Gemini 3.5 ProThe larger and more capable model in the new family is already being used internally by company employees. The new model is expected to be released to the public in [year]. June 2026.

O 3.5 Pro should deliver better performance in tasks that require deep reasoning e understanding long contextsareas where the 3.5 Flash It makes concessions in exchange for speed, although it is still impressive.

Conclusion

In short, it is clear that the Gemini 3.5 Flash This marks an important step forward in Google in the race of artificial intelligence models, especially in the scenario where AI-assisted programming e autonomous agents They have become the main battleground between companies.

With a market share still below rivals such OpenAI e anthropic In this segment, Google is betting on speed, competitive pricing compared to Pro models, and integration with its ecosystem of billions of users to regain ground.

The demonstration with the Antigravity 2.0 The other uses of the Gemini 3.5 Flash are quite impressive, and the model will likely be very helpful for casual parents using the device. Gemini:through the app, or even with Google's most basic plans.

However, the price increase in the API compared to the 2.5 Flash This is a point that developers should carefully evaluate before migrating, especially considering the recent release of... DeepSeek V4 Pro and Flash, with absurdly cheap prices per token considering the size of the models.

And have you tried the new one yet? Gemini 3.5 FlashWhat did you think of the speed and the new features? Tell us in the comments and on our social media! showmetech!

With information: Google Blog | Engadget | Android Authority | The Verge | Financial Times | India Today | 9to5Google | LLM Stats

See also:

Two smart glasses from Samsung and Google will be demonstrated at I/O 2026.

Codex, from OpenAI, can answer emails for you. Here's how.