Google I/O 2025: 주요 AI 업데이트 및 하이라이트

What Google actually announced at Google I/O 2025

Google I/O 2025 was the developer keynote where Google stopped treating generative AI as a side experiment and rebuilt its core products around it. Over two days in May 2025 the company shipped or previewed AI Mode in Search, the Gemini 2.5 model family, Veo 3, Imagen 4, Google Beam, Android XR glasses and a $249.99-a-month subscription tier. This article is a corrected historical record of that event, not a description of Google's products today.

That distinction matters. A lot of what was announced in May 2025 has since shipped, been renamed, been folded into something else, or quietly stopped being mentioned. If you landed here looking for the current state of Gemini, skip to the clearly marked update section near the end, which tracks what happened between the keynote and September 2026.

Each announcement below is described as Google described it at the time, using the figures Google itself published. Where the earlier version of this article got a detail wrong, the correction is stated plainly.

Google I/O 2025 at a glance

The right-hand column is the September 2026 status, so you can see which announcements survived contact with reality.

Announced at I/O 2025	What Google said it was	Status as of September 2026
AI Mode in Search	A conversational, reasoning-led search experience, rolling out to US users	Shipped and expanded well beyond the original US Labs rollout
Gemini 2.5 Pro / Flash	Flagship reasoning model and its faster, cheaper sibling	Superseded by the Gemini 3 series
Deep Think	Experimental enhanced reasoning mode for 2.5 Pro, trusted testers first	Now a named product tier (Gemini 3.1 Deep Think)
Project Astra	Research prototype for a universal multimodal assistant	Still a research prototype; its features feed Gemini Live
Project Mariner	Browser agent research; computer-use skills fed into Agent Mode	No longer a standalone product page; folded into Gemini's agent features
Veo 3	Video generation with natively generated audio	Succeeded by Veo 3.1, which adds 4K output
Imagen 4	Text-to-image model, up to 2K, better typography	Still Google's headline text-to-image model, alongside Nano Banana
Google Beam	3D video calling, the evolution of Project Starline	Enterprise early access via HP Dimension, not a consumer product
Android XR glasses	Prototypes with Gentle Monster and Warby Parker partnerships	Headsets arrived first; glasses remain the slower half of the story
Google AI Ultra	$249.99/month top subscription tier	Repriced and joined by a cheaper Google AI Plus tier

AI Overviews and AI Mode: Search stopped being ten blue links

Google reported that AI Overviews had reached 1.5 billion monthly users across more than 200 countries, and that in its largest markets the feature was driving over a 10% increase in usage of the queries it appeared on. Note the correction: that is a monthly usage figure, not a count of people who said they liked it.

The bigger announcement was AI Mode, a separate tab that runs a query-fan-out technique across many sub-searches and returns a reasoned answer. At I/O 2025 it began rolling out to US users, with Deep Search for longer research tasks, custom charts for data questions, and agentic capabilities for booking event tickets, restaurant tables and local appointments. Search Live, which lets you point a camera at something and keep talking, was promised for that summer.

For publishers, the consequence is that summaries now sit above your listing. Pages that survive extraction are unambiguous about entities and figures and carry meaningful alternative text on every image — an image alt text checker finds the ones you missed faster than reading the HTML yourself.

Gemini 2.5: the model line-up Google showed on stage

Gemini 2.5 was the family Google put at the centre of I/O 2025. It is a thinking model line, meaning it spends compute on multi-step planning before answering, which helps on coding, STEM problems, long documents and large codebases.

Gemini 2.5 Pro was the flagship, which Google described as leading the WebDev Arena and LMArena leaderboards at the time. Gemini 2.5 Flash was the faster, cheaper tier with improved coding and reasoning, made available to everyone in the Gemini app at the keynote. A Flash-Lite tier followed in June 2025.

One correction: Flash-Lite is not a higher-quality model. It is the cheapest and fastest tier, built for high-volume, latency-sensitive work such as classification and translation, and it trades capability to get there.

Google also introduced Deep Think, an experimental enhanced reasoning mode for 2.5 Pro that went to trusted testers via the Gemini API before wider release, and Gemini Diffusion, a research model that generates text using a diffusion process rather than token-by-token. On scale, Google said over 7 million developers were building with Gemini and that it was processing over 480 trillion tokens a month, up roughly fiftyfold year on year, with the Gemini app past 400 million monthly users.

Project Astra: the prototype behind Gemini Live

Project Astra is Google DeepMind's research prototype for a universal assistant that can see, hear and talk in real time. At I/O 2025 it gained native audio output, improved memory and computer control, an accessibility collaboration with Aira, and a conversational tutor prototype for homework.

The key point, then and now, is that Astra is not a product you download. It is the research vehicle whose capabilities graduate into shipping features — camera input, screen sharing and video understanding all reached Gemini Live this way, and Google said they would flow into Search and glasses too.

Another correction: the original version of this article said Gemini Live was powered by Gemini 1.5 Pro. It was not. By I/O 2025 the live experience was running on the 2.5 generation, with Project Astra's native audio and computer-control work behind it.

Project Mariner, not "Project Manier"

The browser agent is Project Mariner, not "Project Manier". It first appeared in December 2024 built on Gemini 2.0, but the I/O 2025 update was a different animal: it could work on several tasks in parallel instead of commandeering your single browser tab, and it added a teach-and-repeat approach where you demonstrate a task once and the agent generalises it.

More importantly, Mariner stopped being just a demo. Its computer-use capabilities were made available through the Gemini API, and they powered Agent Mode, the experimental goal-completion feature Google gave to AI Ultra subscribers in the Gemini app. That handoff is why Mariner's own branding faded — more on that in the update section.

Generative media: Veo 3, Imagen 4, Flow and Lyria

Veo 3 was the headline: a video model that generates sound — dialogue, ambient noise and effects — natively alongside the picture, rather than requiring you to score it afterwards. At launch it went to AI Ultra subscribers in the Gemini app and to Vertex AI. The original article referred to it as "Veo.io"; there is no such product, and the model outputs at 1080p, not an unspecified "high resolution".

Imagen 4 improved fine detail, supported multiple aspect ratios at up to 2K, and got significantly better at spelling and typography inside images. The much-quoted speed claim needs care: Google said a Fast variant would be up to 10x faster than Imagen 3 and was coming soon. Imagen 4 itself was not simply "ten times quicker". All output carries SynthID watermarking, and Google launched a SynthID Detector portal, noting that over 10 billion pieces of content had been watermarked.

Around the models, Google launched Flow, an AI filmmaking tool with character, scene and style controls, plus Lyria 2 for music and Lyria RealTime for interactive generation. If you are drafting scenes before spending render credits, a plain AI image generator is a cheap way to test a look, and an AI video summarizer pulls the beats out of reference footage before you write a prompt.

Workspace, Gmail and shopping

Gmail got personalized smart replies that draw on your Drive context and past emails to match your tone, announced for later in 2025 and requiring your permission. Google Meet gained real-time speech translation that preserves the speaker's voice and expressiveness. Google Vids opened to Pro and Ultra users, and NotebookLM shipped a mobile app.

On shopping, Google connected AI Mode to the Shopping Graph and added virtual try-on, where you upload a single full-length photo and see garments rendered on your own body, plus agentic checkout, which watches a price and buys at your target with your confirmation. The original article's claim that this "tracks discount offers" is roughly right, but the mechanism is a price-tracking agent tied to a checkout flow, not a coupon finder.

Beyond software: Google Beam and Android XR

Google Beam is the renamed, productised Project Starline: a light-field 3D video system that makes a remote person appear life-sized without a headset, built with HP and with Zoom interoperability announced. Android XR covered Samsung's Project Moohan headset and prototype glasses shown with language translation, messaging and directions, with eyewear partnerships announced for Gentle Monster and Warby Parker.

What Google I/O 2025 cost you

The subscription restructure was one of the most consequential announcements. At the keynote Google offered:

Google AI Pro at $19.99 a month, with higher rate limits, Flow access and Gemini in Chrome on desktop.
Google AI Ultra at $249.99 a month, discounted 50% for the first three months for new subscribers, including Veo 3, Agent Mode, 30TB of storage and YouTube Premium.
A free yearly upgrade for college students in the US, Brazil, Indonesia, Japan and the UK.

Update: what has happened since Google I/O 2025

This section is written in September 2026 and describes the state of play now, more than a year after the keynote. Treat everything above as the historical record and everything here as the correction layer.

The models moved on twice over

Gemini 2.5 is no longer the current generation. Google's line-up has moved to the Gemini 3 series, with Gemini 3.1 Pro as the general-purpose flagship and Gemini 3.1 Deep Think as a named tier aimed at science, research and engineering. Deep Think went from an experimental toggle for trusted testers to a product you can buy. The Flash line has iterated fastest, and now includes a Flash-Lite tier and a security-focused Flash Cyber variant, alongside specialised models such as Gemini Omni for multimodal creation and Gemini Robotics.

Names that changed or quietly disappeared

Project Mariner no longer has its own product page at Google DeepMind — its agentic browsing work now surfaces inside the Gemini app rather than under the Mariner brand, and the "Agent Mode" label from I/O 2025 has itself been superseded by newer agent features in the app. Project Astra, by contrast, is still explicitly a research prototype with a tester waitlist, exactly as it was in 2025; it never became a shipping product, and Google's own page still frames it as the source of features that graduate into Gemini Live.

Generative media kept moving

Veo 3 has been succeeded by Veo 3.1, which generates audio natively as before but now outputs 4K as well as 1080p. Imagen 4 is still Google's headline text-to-image model, but it has been joined by the Nano Banana image line for generation and editing with strong subject consistency and text rendering. Flow survived and grew: it is now positioned as a creative studio driven by Gemini Omni, Nano Banana and Veo 3.1 together, with an agent that helps plan and refine a project rather than just accepting prompts.

The pricing was rebuilt

The $249.99 Ultra tier from the keynote no longer stands alone at that price. Google now sells three consumer tiers — Google AI Plus at $4.99 a month, Google AI Pro still at $19.99, and Google AI Ultra in a $99.99 to $199.99 range depending on the plan — with generous free daily allowances in tools like Flow. Pro is the one price that has held steady since I/O 2025.

Hardware landed unevenly

Google Beam is real but narrow: it ships as HP Dimension for enterprise customers under limited early access, not as something you can buy for a home office. Android XR headsets arrived ahead of the eyewear, and the consumer smart glasses previewed on stage in 2025 have been the slower half of that programme. Meanwhile Google I/O 2026 has been and gone, which is a useful reminder that any "latest Google AI" article carries an expiry date.

How to check whether an I/O 2025 feature still exists

Keynote recaps rot quickly. Before you rely on anything here, or in any other summary, run this check:

Find the vendor's own page for the product, not a news write-up. If Google DeepMind or blog.google no longer has a page for it, the feature has almost certainly been folded into something else.
Check whether the URL redirects. A product page that now bounces to a company homepage is the clearest signal that a name has been retired.
Confirm the model version on the official model page rather than trusting a number quoted in an article. Version numbers in this space change every few months.
Re-check the price on the vendor's plans page in your own region, since AI subscription pricing has been repeatedly restructured and varies by country.
Date-stamp what you write. If you are summarising a keynote, say when you wrote it, and use an AI text summarizer on the vendor's own recap post rather than paraphrasing second-hand coverage.

If you publish in more than one language, run the finished English version through an AI translator only after the facts are locked, and consider checking your draft with an AI content detector if editorial policy requires disclosure. For planning which of these topics is still worth targeting, an AI keyword generator will show you whether readers are still searching for the 2025 product names or have moved on to the current ones.

Frequently asked questions

When was Google I/O 2025 and where was it held?

Google I/O 2025 took place in May 2025 at the Shoreline Amphitheatre in Mountain View, California, with the main developer keynote streamed free worldwide. Google publishes a full written recap of every announcement on its own blog, which is the source this article follows.

Is Gemini 2.5 Pro still Google's best model?

No. Gemini 2.5 Pro was the flagship announced at I/O 2025, but Google's line-up has since moved to the Gemini 3 series. As of September 2026 the flagship reasoning tiers are Gemini 3.1 Pro and Gemini 3.1 Deep Think, with faster Flash and Flash-Lite tiers below them.

What happened to Project Mariner?

Project Mariner was DeepMind's browser-agent research. Its computer-use capabilities were released through the Gemini API and powered Agent Mode in the Gemini app, and the standalone Mariner branding has since disappeared from Google DeepMind's site. The capability lives on inside Gemini's agent features rather than as a separate product.

Can I use Veo 3 for free?

Veo 3 launched at I/O 2025 for Google AI Ultra subscribers and on Vertex AI, so it was not free at announcement. Video generation is now handled by Veo 3.1 and is metered by credits, with a limited free daily allowance in Flow and larger allowances on the paid Google AI plans. Check Google's current plan page for your country before assuming a quota.

Did Google I/O 2025 announce anything that never shipped?

Several things shipped far more narrowly than the keynote implied. Google Beam exists only as an enterprise device from HP under limited early access rather than a general product, the smart glasses shown with Gentle Monster and Warby Parker lagged well behind the headsets, and Project Astra is still a waitlisted research prototype rather than an app.

Are AI Overviews and AI Mode the same thing?

No. AI Overviews are the generated summaries that appear above ordinary results for some queries, and Google said they reached 1.5 billion monthly users in over 200 countries by I/O 2025. AI Mode is a separate tab you switch into for a full conversational search session with follow-ups, deeper research and agentic booking.

The takeaway from Google I/O 2025

Read a year later, Google I/O 2025 looks less like a list of features than a commitment: Search, Workspace, Android and DeepMind's research pipeline were all pointed at the same model family, and the subscription tiers announced that week set the pattern for how Google has sold AI ever since.

What aged badly is the specifics. Model numbers rolled over twice, one flagship price fell by hundreds of dollars, and at least two named projects stopped being names. That is normal for this field, and exactly why a keynote recap should be dated, sourced to the vendor's own posts, and revisited rather than left to go stale.