Google DeepMind
Google DeepMind is Google's AI research lab, born from the merger of DeepMind and Google Brain. It builds and publishes the Gemini, Veo, Imagen, Lyria, Genie and Gemma model families, alongside scientific systems such as AlphaFold.
What is Google DeepMind?
Google DeepMind is Google's artificial intelligence laboratory, formed by bringing together two research groups: DeepMind, started in 2010 with an interdisciplinary approach spanning machine learning, neuroscience, engineering, mathematics and simulation, and Google Brain, started in 2011 at X, the moonshot factory. The combined team is led by co-founder and CEO Demis Hassabis.
Its record explains the ambition. DeepMind produced DQN, which learned forty-nine Atari games from raw pixels; AlphaGo, the first program to beat a Go world champion, in 2015; AlphaZero and MuZero; AlphaStar, the first system to defeat a top professional at StarCraft II, in 2019; WaveNet, which became the voice of Google Assistant; and AlphaFold, which predicted 3D protein structures in 2020. Google Brain invented the Transformer architecture in 2017, open-sourced BERT, JAX and TensorFlow, and built the LaMDA and PaLM model families.
The site organises today's output into four groups. Gemini covers frontier text and agent models, from 3.6 Flash to 3.1 Pro, plus Gemini Omni, Gemini Image, Gemini Audio and Gemini Embedding. Specialised models gather Veo for cinematic video with audio, Imagen for text-to-image and Lyria for music. World models and physical AI hold Genie 3, which generates explorable interactive environments, and Gemini Robotics, presented as an intelligence layer for any kind of robot, with an on-device variant. Gemma is the open model family.
A separate science strand applies the same methods to research: AlphaFold, cited as used by more than three million researchers, AlphaGenome, WeatherNext for weather and cyclone forecasting, AlphaEarth for planetary mapping, AlphaEvolve for algorithm design, and Co-Scientist. SynthID embeds imperceptible watermarks into generated images, video, audio and text, designed to survive cropping, filters, frame-rate changes, lossy compression and added noise.
Note what the site is not: a shop. It publishes no prices, no plans and no contact page. The models are used elsewhere, on the Gemini app, Google AI Studio, the Gemini API, Google Antigravity and Google Flow.
What it does
- Generate text, code and agentic workflows with the Gemini model family
- Create and edit images with Gemini Image, also known as Nano Banana, and with Imagen
- Produce cinematic video with synchronised audio through Veo and Gemini Omni
- Compose high-fidelity music and speech with Lyria and Gemini Audio
- Generate and explore interactive 3D worlds with Genie 3
- Run open models locally with the Gemma family
- Watermark and identify AI-generated images, video, audio and text with SynthID
When to use Google DeepMind / When not to
A quick filter to help you decide if Google DeepMind is the right fit.
When to use Google DeepMind
- Developers building on the Gemini API or running the open Gemma models locally
- AI and machine-learning researchers who need model cards, evaluations and the lab's published papers
- Scientists in biology, genomics, weather and earth observation working with AlphaFold, AlphaGenome, WeatherNext or AlphaEarth
- Creative professionals generating video, images, music or interactive worlds with Veo, Imagen, Lyria and Genie
- Robotics and embedded teams evaluating Gemini Robotics, including its on-device variant
When not to use Google DeepMind
- Buyers who need a published price list: the site sells nothing and shows no plans
- Anyone looking for a support channel, as there is no contact page and no email address on the domain
- Teams needing DeepMind-specific legal terms, since the applicable documents are Google's general terms and privacy policy
- Mobile-first users expecting a dedicated DeepMind app: no App Store or Google Play listing is offered
- Procurement or compliance teams requiring a DPA, a subprocessor list or a documented training opt-out
How to use Google DeepMind
A typical end-to-end flow, from setup to results.
- Open deepmind.google and browse the Models section to see the current catalogue by modality
- Read the model card and the Evals page for the model you are considering
- Pick an access surface from the 'Try it in' block shown on each model page
- Sign in with a Google account, which the product surfaces require even though the showcase site does not
- Try a model conversationally in the Gemini app, or interactively in Google AI Studio
- Follow the official prompt guide for your modality, whether video, image, music or worlds
- For programmatic use, read the Gemini API documentation on ai.google.dev and obtain an API key
- For open models, get Gemma through AI Studio or the Run Gemma documentation and run it locally
- Consult the research and breakthroughs pages for methodology, benchmarks and publications
- Subscribe to the newsletter in the footer to follow new model releases
Pros & Cons
Pros
- Complete multimodal coverage from a single lab: text, image, video, audio, music, 3D worlds and robotics
- Open models available alongside proprietary ones, so work need not be locked to a hosted API
- Benchmarks published with methodology and explicit comparison against named competing models
- Model cards and a dedicated evaluations page make model behaviour auditable
- Deep research record, with 262 publications listed in the site's own sitemap
- Named, formalised safety governance: AI Principles, Responsibility and Safety Council, AGI Safety Council and the Frontier Safety Framework
- SynthID provides provenance for generated content across all four modalities
Cons
- No pricing page and no plans: the only figures on the site are API token prices inside a benchmark table
- No contact page and no email address anywhere on the domain
- Legal documents are Google's general terms and privacy policy rather than DeepMind-specific ones, and their content varies by the reader's region
- No DPA, no subprocessor list and no documented training opt-out in the pages collected
- Hosting countries unspecified, since Google states data may be processed in any country of the world
- No minimum age given as a figure; the terms defer to the age required for a Google account
- Actual use is dispersed across several Google surfaces, none of them on deepmind.google itself
Pricing & Plans
Google DeepMind publishes no pricing page and no plans on its website. Free access to the models exists in practice through the Gemini app and Google AI Studio, though the site never describes this as a free plan, and no time-limited free trial is mentioned. The only figures shown are API token prices in a benchmark comparison table on the Gemini model page: 1.50 USD per million input tokens and 7.50 USD per million output tokens for Gemini 3.6 Flash, 1.50 and 9.00 USD for Gemini 3.5 Flash, and 2.00 and 12.00 USD for Gemini 3.1 Pro. These amounts are presented for comparison rather than as a commercial offer, and should be confirmed in the API documentation before any commitment.
Data, GDPR & hosting
A consolidated view of how Google DeepMind handles your data.
GDPR overview
Google's privacy policy carries a dedicated European requirements section and states outright that it covers Google's compliance with EU and UK data protection law. It lists the rights available, namely access, update, deletion, restriction, objection and portability, and enumerates five legal bases: contract, legitimate interests, vital interests, legal obligation and consent. Two controllers are named: Google Ireland Limited, Gordon House, Barrow Street, Dublin 4, Ireland for the EEA and Switzerland, and Google LLC, Mountain View, California for the United Kingdom. The data protection office is reachable through a support form rather than a published email address. No Article 27 representative is designated, which is consistent, since the EEA controller is established inside the Union. The Swiss representatives named in the policy answer to Article 14 of Swiss law, not the GDPR.
Who owns the data?
Google's Terms of Service state that your content remains yours: you keep any intellectual property rights you hold in it. Google, for its part, retains the intellectual property rights it holds in the services themselves, and forbids removing or altering its branding, logos and legal notices. Feedback is treated differently: if you send suggestions, Google may act on them without any obligation towards you. Note that these are Google's general terms rather than a DeepMind-specific contract, and the contracting entity changes with your location, being Google Ireland Limited in the EEA and Switzerland and Google LLC elsewhere.
Reuse rights
Users keep their own content and may reuse it freely, since the terms transfer no ownership to Google. What Google does with the information it collects is set out in its privacy policy, effective 2 April 2026: content you create or upload, usage signals from devices, browsers and activity, location data, and information gathered from publicly accessible sources. Google states that publicly available material helps train its AI models, and names Google Ireland Limited as the controller responsible for processing information to train those models for services delivered in the EEA and Switzerland. Personal information is not shared outside Google except with your consent, with domain administrators, with affiliates and service providers acting on Google's instructions, or for legal reasons. Whether customer content itself feeds model training is not settled in these documents, and the Gemini Apps carry a separate privacy notice that this record does not cover.
Data retention & training
Hosting summary
Google DeepMind's website states no hosting jurisdiction of its own, and the applicable document is Google's privacy policy. It says Google maintains servers around the world and that information may be processed on servers located outside the country where you live and, in its own wording, possibly in any country of the world. Affiliates and service providers receiving that information may likewise be located in, or process it from, any country. Google says the protections described in the policy apply regardless of where processing happens, and that it complies with certain legal frameworks governing data transfers, without listing them on the page collected. For users in the EEA and Switzerland the controller is Google Ireland Limited in Dublin; for the United Kingdom it is Google LLC in Mountain View, California. No country list and no regional commitment such as EU-only or UK-only hosting is published, which is why neither hosting countries nor hosting regions could be recorded for this tool.
Things to keep in mind
Risks and trade-offs to weigh before adopting Google DeepMind.
- Over-reliance on generative models can erode the very skills they seem to replace, from writing to reasoning through a problem, especially when output is accepted without review
- The applicable terms are Google's general ones and the contracting entity changes with your location, so the rights you actually hold depend on where you are
- Whether customer content feeds model training is not settled in the published documents, and no opt-out mechanism is documented
- Data may be processed in any country in the world, with no hosting jurisdiction guaranteed
- Generated video, images, audio and text are realistic enough to mislead; SynthID watermarking helps, but only where it is applied and actually checked
- Benchmark scores and API prices shown on the site are comparisons published by the vendor, not contractual commitments
- Capabilities move quickly, with variants in preview or announced as coming soon, so behaviour observed today may not hold tomorrow
Setup & Integrations
Technical difficulty
Browsing the site requires nothing at all: no account, no installation. Trying a model is almost as simple, since the Gemini app and Google AI Studio need only a Google account, and official prompt guides cover video, image, music and world generation for newcomers. Difficulty rises with programmatic use: the Gemini API assumes developer skills, an API key and time with the documentation. Running the open Gemma models locally is the most demanding path, requiring suitable hardware. In short: trivial to explore, easy to try, moderate to integrate.
Deployment
Behind Google DeepMind
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Google DeepMind.
Frequently asked questions
What is Google DeepMind?
Which models does it produce?
Where can I actually try these models?
Is there an API?
How much does it cost?
Are there open models?
How can AI-generated content from these models be identified?
Which privacy policy and terms apply?
Who is the data controller in Europe?
How do I contact Google DeepMind?
Should you pick Google DeepMind?
Google DeepMind's website is the shop window of a research laboratory, not a product you buy. Read with that in mind, it is unusually substantial: a full catalogue of models spanning text, image, video, audio, music, interactive worlds and robotics; an open family, Gemma, sitting beside the proprietary ones; benchmarks published with their methodology and compared against named competitors; model cards; a dedicated evaluations page; and 262 research publications listed in the site's own sitemap. The science strand is the most distinctive part, with AlphaFold cited as serving more than three million researchers, and AlphaGenome, WeatherNext, AlphaEarth and AlphaEvolve extending the same approach to genomics, weather, planetary mapping and algorithm design. Safety governance is named rather than gestured at, from the AI Principles to the Frontier Safety Framework, and SynthID gives generated content a provenance signal across all four modalities.
The gaps are equally clear and worth stating plainly. There is no pricing page, no plan, no contact page and no email address on the domain. The legal documents that govern use are Google's general terms and privacy policy, whose content and contracting entity change with the reader's region, and none of the pages collected mention a DPA, a subprocessor list or a way to exclude data from model training. Hosting is described only as servers around the world. Nothing here is actually used on deepmind.google either: the models live on the Gemini app, Google AI Studio, the Gemini API, Antigravity and Flow.
Treat the site as the authoritative reference on what these models are and how they were evaluated, then go elsewhere in Google's estate to use them and to find out what they cost.
- Choosing a selection results in a full page refresh.
- Opens in a new window.