Google AI: Research, Gemini, Search and Alphabet

Google’s AI history includes transformers, TensorFlow, BERT, Gemini and AI Overviews. It remains the operating company beneath Alphabet.

kindCompany
foundedSep 1998
part ofAlphabet

Latest

Updated

Research, models and distribution

Google is an operating company within Alphabet. The group's annual filing dates Google's incorporation to September 1998; the October 2015 reorganization put it under the new holding company. Search, cloud services and other Google products supply routes through which AI research reaches users.

Google researchers' 2017 Transformer paper introduced an attention-based sequence architecture that became important to language-model development. That contribution does not mean every later major AI model uses the same architecture. TensorFlow's 2015 software release and BERT's 2018 pre-training research are separate contributions to how developers train and use models.

In April 2023, Google combined DeepMind and the Brain team into Google DeepMind. This identifies a particular model-research organization within the wider company; it does not make Google, historical DeepMind and every Google Research project interchangeable actors.

Putting the models into products

Gemini's December 2023 introduction connected model development with Google's consumer and developer products. AI Overviews in May 2024 put generated answers into Search, a different product milestone from releasing the underlying model.

The January 2026 agreement with Apple extended Google's technology into another device ecosystem. The parties said Apple's next foundation models would be based on Gemini and Google cloud technology while continuing to run through Apple's on-device and Private Cloud Compute architecture. Licensing technology did not transfer ownership of Apple to Google.

Sources

  1. [1]
    Alphabet Inc. Form 10-K for fiscal 2025

    Alphabet Inc. · Feb 4, 2026

  2. [2]
    Attention Is All You Need

    arXiv · Jun 12, 2017

  3. [3]
  4. [4]
  5. [5]
  6. [6]
  7. [7]
  8. [8]
  9. [9]

Google history

2026

│
Google opens immersive explorations in Gemini

Google made an interactive visual-topic library available in Gemini on web and mobile, including education accounts and personal Google accounts. Six languages were supported, with an English experience for other languages; in-chat access and narration remained forthcoming.

Google
Product
│
Google brings Gemini Personal Intelligence to Android Auto

Google began rolling out opt-in access to personal Google app data through Gemini in Android Auto, with cars running Google built-in to follow in the coming months. The rollout covered eligible users aged 18 and older with personal accounts outside the EEA, Switzerland, the UK and Nigeria; Workspace business and education accounts were excluded.

Google
Product
│
Google releases Gemini 3.8 Live with Live Avatar

Google made Live Avatar available in Gemini Enterprise, pairing Gemini 3.8 Live speech with generated video personas that process audio and visual input while calling tools. Preset avatars are available; creating one from a reference image requires enterprise allowlisting.

Google
Model
│
Google releases Gemini 3.8 Live and 3.8 Live Extended Thinking

Two speech-to-speech models: a cheaper Live built for scale, and an Extended Thinking version that reasons while it talks and runs tool calls in the background. Artificial Analysis ranked Extended Thinking first on its speech-to-speech index, 1.1 points ahead of OpenAI's GPT-Live-1.

Google
Model
■
Apple ships Siri AI in beta, on models built with Google's Gemini

With Apple

iOS 27 and Apple's other 2027 software releases brought the rebuilt Siri to supported devices in English, with daily limits on server-side features and paid expanded access promised later. It launched without EU availability on iPhone, iPad and Watch, and without China.

Apple
Product
│
More than a hundred companies sign a call for collective action on cyber defense

With OpenAI, Anthropic, Microsoft, Amazon, Cloudflare, Hugging Face, Oracle, Perplexity

OpenAI hosted the letter and Anthropic, Google, Microsoft, AWS, Cloudflare, CrowdStrike, Visa and Mastercard signed it, the day after the Hugging Face postmortem. The premise is a window: models now help defenders faster than attackers, and the signatories expect it to close within months.

OpenAI
Culture
■
Marvell gives Google a warrant worth $12.2 billion in a custom chip deal

With Marvell

Marvell issued Google a warrant for nearly 59 million shares, worth $12.2 billion if fully exercised, most of which only vests as Google buys custom AI inference, storage and networking silicon from Marvell through fiscal 2033. The arrangement would make Google one of Marvell's largest shareholders without either company disclosing a dollar figure for the chip orders behind it.

U.S. Securities and Exchange Commission
Deal
■
The Gemini app passes a billion monthly users

Google's fourteenth product to cross a billion and the fastest to get there, running at 150 million images a day with 63 percent of users talking to it out loud. It went from 950 million to a billion in under a month.

Google
Culture
■
Hassabis moves to chair and Jeff Dean leaves Google

With Google DeepMind

Koray Kavukcuoglu took over Gemini, the app, and frontier research as senior vice president; Demis Hassabis became chair of Google DeepMind and chief scientist of Alphabet; and Jeff Dean left after 27 years to start a public benefit corporation with Sanjay Ghemawat.

Google
Company
│
Google limits Meta's access to its Gemini models

With Meta

Meta had been buying Gemini capacity for safety filtering, coding assistants and internal tools; Google could not fill the order and capped it, delaying some of Meta's own AI projects. Even a company building its own frontier models still needed someone else's compute.

CNBC
Deal
■
Anthropic expands its Google and Broadcom TPU deal

With Anthropic, Broadcom

Anthropic signed an agreement for multiple gigawatts of next-generation Google TPU capacity expected to come online beginning in 2027. Broadcom said about 3.5 gigawatts would be made available through it, subject to Anthropic's continued commercial success.

Anthropic
Deal
│
Google releases Gemini 3.1 Flash Live

The audio model behind Gemini Live and Search Live, tuned for lower latency, for hearing pitch and pace, and for holding a conversation in over 200 countries. Everything it says carries a SynthID watermark.

Google
Model
│
Google releases Nano Banana 2

Formally Gemini 3.1 Flash Image, it puts the Pro line's world knowledge, text rendering and subject consistency into something fast enough to sit inside a loop. Output goes to 4K.

Google
Model
■
Google releases Gemini 3.1 Pro

Google released Gemini 3.1 Pro in preview across its API, enterprise products, the Gemini app and NotebookLM, with consumer access gated to Pro and Ultra plans. Its 77.1 percent ARC-AGI-2 result used the Thinking High setting and measured abstract puzzle solving, not general task success.

Google
Model
●
Apple bases its next foundation models on Gemini

With Apple

Apple and Google entered a multi-year collaboration under which the next Apple Foundation Models would be based on Gemini models and Google cloud technology. Apple said its intelligence features would still run on its devices and Private Cloud Compute.

Google
Company

2025

│
Google gives a model a memory that grows

Titans replaces the fixed-size vector an RNN carries with a deep network that learns what to keep, and MIRAS is the framework describing the design space around it. The target is context lengths a transformer cannot afford to attend over.

Google
Research
■
OpenAI calls a code red over Gemini

With OpenAI

Two weeks after Gemini 3 shipped, Altman told staff ChatGPT was at a critical time and pushed the advertising plans back to make room for fixing it. The memo warned of temporary economic headwinds and rough vibes.

Fortune
Company
│
Google puts Gemini 3 Pro behind an image model

Nano Banana Pro, formally Gemini 3 Pro Image, renders legible text in several languages, blends up to 14 images while holding five people consistent, and outputs at 4K. It landed in the Gemini app, Ads, Workspace, Vertex and NotebookLM on the same day.

Google
Model
│
Google ships a model that drives a browser

Gemini 2.5 Computer Use clicks, types and scrolls its way through web interfaces, built on Gemini 2.5 Pro and aimed at browsers first with mobile still emerging. Google claims the lead on Online-Mind2Web, WebVoyager and AndroidWorld.

Google
Model
│
Nano Banana makes image editing a conversation

Google's image model spent weeks being tested anonymously on LMArena under a codename, and the codename is what stuck. Keeping a face consistent across a dozen edits was the feature that made it spread.

Google
Model
│
Google puts an on-device Gemini Nano into every Pixel 10

Magic Cue drafts replies from on-screen context without a network call, Gemini Live can see through the camera, and the Pixel Watch and Buds got their own Gemini access. The on-device model is the part that works without a data connection.

Google
Product
│
Veo 3 generates the soundtrack with the video

Dialogue, footsteps, and street noise produced in the same pass as the pictures, wrapped in a filmmaking tool called Flow. Silence had been the reliable tell for AI video, and this removed it.

Google
Model
│
Google's Jules coding agent enters public beta

Jules clones a codebase into an isolated Google Cloud virtual machine, plans a fix or feature using Gemini 2.5 Pro, and opens a pull request for review. It launched free with usage limits, aimed squarely at OpenAI's Codex and GitHub's Copilot agent.

Google
Product
│
Google introduces the $250-a-month AI Ultra subscription

The top-tier plan bundles the highest Gemini usage limits, early access to Veo 3 and the Flow filmmaking tool, an agent called Project Mariner that juggles ten browser tasks at once, and 30TB of storage. Google's existing paid tier is renamed AI Pro.

Google
Product
│
Google unveils Ironwood, its seventh-generation TPU

A chip built for inference rather than training, which Google calls its most powerful and power-efficient TPU yet. Announced at Google Cloud Next as part of a broader pitch that inference, not training, is now the harder problem.

Google
Hardware
│
Google introduces the Agent2Agent protocol for AI agents

An open specification so agents built on different frameworks and by different vendors can advertise their capabilities and hand off tasks to each other. Google positioned it as a complement to Anthropic's Model Context Protocol rather than a rival.

Google
Open source
■
Google agrees to buy Wiz for $32 billion

With Wiz

The largest acquisition in Alphabet's history, all cash, for a cloud security company five years old. An earlier run at $23 billion had fallen apart the previous summer.

Google
Acquisition
│
Gemma 3 fits a capable open model on one GPU

Four sizes from 1B to 27B, vision above the smallest, and up to 128,000 tokens of context, with Google claiming the 27B beat models many times its size. The hardware requirement was the interesting number, not the benchmark.

Google
Model

2024

│
Second lawsuit accuses Character.AI of harming minors

With Character.AI

Families of a Texas teenager and a nine year old girl sued Character.AI, Google and the company's founders in federal court, alleging chatbots encouraged self harm and exposed the children to sexual content.

NPR
Culture
│
Mother sues Character.AI over teen son's death by suicide

With Character.AI

Megan Garcia alleged that Character.AI's product design, including sexualized conversations with her 14 year old son Sewell Setzer, contributed to his death, and named Google and Alphabet as co-defendants. It became a rallying point for concerns over AI companion safety.

NBC News
Culture
│
Frontier labs sign the Seoul safety commitments

With OpenAI, Anthropic, Meta, Microsoft, Amazon, Mistral AI, xAI, Cohere, G42, IBM, Inflection AI, Naver, Samsung Electronics, Technology Innovation Institute, Zhipu AI

At the UK and South Korea's Seoul summit, sixteen frontier labs including OpenAI, Anthropic, Google, Meta and Zhipu AI committed to publish safety frameworks with risk thresholds and to hold back models that breach them. Voluntary, and the signatory list kept growing afterwards.

Department for Science, Innovation and Technology
Policy
│
Gemini 1.5 Flash and Project Astra debut at Google I/O

With Google DeepMind

Google's answer to GPT-4o's launch day: Gemini 1.5 Flash for cheap high-volume work, a two million token context window, and Project Astra, a camera-and-voice assistant demo. Flash became the workhorse; Astra stayed a demo.

Google
Model

2023

■
Google answers with Gemini

Three model sizes, native multimodality, and a launch video that turned out to have been edited for pace. The model was real and the credibility cost was real too.

Google
Model

2022

2018

■
BERT learns language from both directions

Google trained BERT to recover hidden words using context on both sides, then adapted it to tasks such as answering questions from a passage. The paper reported leading results on eleven language tasks; code and pretrained models followed later.

arXiv
Research
│
Google Duplex books a haircut over the phone

An assistant called a salon, said "mm-hmm" in the right places, and made an appointment without the person on the line appearing to notice. The demo drew a week of argument about disclosure and Google conceded that the calls would identify themselves.

Google Research
Product

2017

●
Attention Is All You Need introduces the transformer

The Transformer used attention to connect words directly, making translation-model training more parallel. Its encoder and decoder became starting points for BERT and GPT, which adapted the architecture to learn from unlabeled text.

arXiv
Research

2016

│
Technology companies establish Partnership on AI

With Partnership on AI, Amazon, Meta, DeepMind, IBM, Microsoft

Amazon, Facebook, Google, DeepMind, IBM and Microsoft established a nonprofit forum for AI research and practices. The Partnership on AI was publicly announced in New York on September 28.

DeepMind
Company

2015

■
Google open sources TensorFlow

Google released its internal machine learning framework and set the default toolchain for the next five years. Owning the ecosystem turned out to be worth considerably more than keeping the code.

Google Research
Open source

2014

●
Google acquires DeepMind

With DeepMind

Google paid a reported £400 million for a London lab with about fifty employees and no product. The price looked absurd at the time and then, gradually, did not.

The Guardian
Acquisition

2013

■
word2vec turns words into arithmetic

Mikolov’s group at Google introduced efficient methods for training word vectors on billions of words. The learned vectors supported analogies, including the relation between man and woman and between king and queen.

arXiv
Research

2012

■
Google finds a cat neuron in YouTube

A network trained on unlabeled YouTube frames developed a feature responsive to cat faces. The experiment used 16,000 cores and about a billion connections to study unsupervised learning at scale.

Google
Research