Scientists today face challenges of extraordinary scale and complexity. From shaping and simulating the convoluted dynamics of fusion plasma, through exploring the enormous search space for novel materials, to understanding exabytes of data flowing from the most advanced experimental facilities in the world. The demands placed on state-of-the-art research are unprecedented. Frontier AI can aid address these challenges while accelerating scientific breakthroughs.

In December, we shared our commitment to the White House Mission Genesis — a nationwide effort to leverage artificial intelligence and double the pace of American scientific discovery within a decade. Google DeepMind (GDM) has since announced an early access program that brings AI-powered scientific tools to all 17 Department of Energy (DOE) national laboratories and Google’s public sector shared, like Gemini for the government could serve as an artificial intelligence framework for the DOE.

Today at the DOE Genesis Mission Summit 2026, we are expanding that scope by committing $40 million in AI tokens and cloud credits to researchers supporting the Genesis Mission.

Frontier AI tools for scientific discovery

As part of this expanded commitment, we will first provide DOE Genesis Mission winners with in-kind access to GDM’s pioneering portfolio of artificial intelligence for science, including:

• AlphaEvolve – Gemini-based coding and discovery agent for designing advanced algorithms.
• AlphaFold 3 — a model for predicting the structure and interactions of proteins and other biomolecules.
• AlphaGenome — a tool for understanding how variation in DNA, including the non-coding genome, shapes biology and disease.
• WeatherNext — a state-of-the-art family of AI weather forecast models for mapping weather conditions.
• AlphaEarth Foundations — the foundational artificial intelligence model that enables us to map and understand our planet in unprecedented detail.

Second, we will make Gemini available for government jobs and tokens for one year to tens of thousands of users across DOE National Laboratories operations, research and management teams. This secure platform supports the full scope of work, from the test bench to the administration of specialized user facilities serving the entire scientific community, providing one secure foundation on which DOE’s mission can rely.

Artificial intelligence for scientific tools in action across the entire laboratory ecosystem

While we still have a long way to go, the practical impact of the Genesis mission is already starting to come to life across the lab ecosystem.

On Pacific Northwest National Laboratory (PNNL)senior scientist Dr. Henry Kvinge uses AlphaEvolve to map enormous mathematical systems that are too convoluted for humans to study by hand. Artificial intelligence automatically discovers hidden connections, allowing for rapid discoveries that would normally take researchers years to find.

“Modern mathematics is based on abstractions, but combinatorics offers concrete models that make it easier to understand complex geometry and algebra. We found that systems like AlphaEvolve are ideal for such explorations,” said Dr. Kvinge. “By leveraging the LLM’s extensive mathematical knowledge, we can automate the study of countless angles. We are still experimenting, but the discoveries are already shaping our future research.”

On Rocky Mountain National Laboratory (NLR)researchers are using Gemini to fundamentally change the way they interact with physical laboratory equipment. Dr. Steven R. Spurgeon, senior materials data scientist at NLR, leads a pioneering program in autonomous materials discovery.

“Our collaboration has allowed us to build autonomous capacity to conduct experiments,” said Dr. Spurgeon. “By implementing Gemini on our instruments, we reduced the microscope calibration time from over 90 minutes to approximately 13 minutes (eight times faster) and reduced the number of manual steps needed to sharpen the image from as many as 50 to two. This is time and attention that we have given to the science itself, enabling truly autonomous workflows that observe, reason and make decisions in real time. This has helped us explore areas of the materials design space that we simply could not reach through human operation.” “manual”.

Powering American Innovation

Mission Genesis represents an opportunity to transform research and science across America. By providing access to advanced artificial intelligence tools, we want to aid scientists accelerate breakthroughs in key energy, security and science challenges. To learn more about how AI capabilities can support your research initiatives, join us at an upcoming event Google Public Sector Summit in October.

• 3.6 Flash: Our workhorse model that delivers better coding, knowledge work, and multimodal efficiency. According to Artificial Analysis Indexreduces output token usage by 17% compared to Flash 3.5, and in some benchmarks such as DeepSWE, by Data curvewe see up to 65%, all at a lower cost per output token.
• 3.5 Flash version: Our fastest and most cost-effective Class 3.5 model, delivering 350 output tokens per second according to the Artificial Intelligence Index, also significantly outperforming previous generations of Flash-Lite in agentic workflows.
• 3.5 Flash Cyber ​​in CodeMender: Effective cybersecurity applications require careful coordination of the model with the agent infrastructure. We are introducing a combination of a novel, high-performance, specialized cybersecurity-focused model combined with our CodeMender code security agent that delivers competitive performance at the edge.

In addition to today’s releases, Gemini 3.5 Pro is currently in testing with partners and we plan to make it widely available as soon as it is ready. At the same time, our team is already focusing on building novel generation models. We have launched our most ambitious pre-training for Gemini 4 to date and we are thrilled with the progress.

3.6 Flash: More capable and better quality than 3.5 Flash

Gemini 3.6 Flash is based directly on developer and customer feedback on 3.5 Flash. 3.6 Flash not only advances coding and knowledge work, but it does so while significantly improving token performance. For example, in the Artificial Intelligence Index, we see that 3.6 Flash uses 17% fewer output tokens than 3.5 Flash. Implementing multi-step workflows also requires fewer reasoning steps and tool calls.

This increased performance is also combined with a lower price than 3.5 Flash. With $1.50/1M input tokens and $7.50/1M output tokens, Flash 3.6 reduces the overall cost of an agent job, making it more cost-effective to build and run agents.

The global biosecurity landscape is changing rapidly. Changes to natural ecosystems, global travel and the potential misuse of AI require greater vigilance, but AI is also a key tool in our response. We need pioneering AI models and the scientific advances they enable to address these challenges and assist make society more resilient to events such as future epidemics.

Today Google DeepMind i Isomorphic laboratories they share ours a common approach to biological resilience.

Our work is twofold – to prevent our models from being misused by threat actors, and to empower governments, scientists, biosecurity experts and our teams to employ these technologies to build a more resilient world.

Over the last 12 months, we have established more than 15 partnerships with governments, biosecurity organizations and research groups to prevent threat actors from misusing our models, quickly detect modern disease outbreaks, and respond quickly and effectively.

As part of our biological resistance program

We believe that society must leverage the growing capabilities of artificial intelligence to combat infectious diseases and prepare for future epidemics.

Breakthroughs like Google DeepMind’s AlphaFold, which maps the 3D structures of almost all known proteins; Powered by Isomorphic Labs AI Drug design engine (IsoDDE), which provides the real-world accuracy required to navigate modern biological systems with unprecedented precision; and AlphaGenome, which shed airy on genome function, radically change the balance. Instead of simply responding to natural epidemics or security threats, we can now employ these smart systems to assist scientists design proactive defenses, accelerate therapeutic discovery, and protect the global health ecosystem with greater speed and precision.

With this in mind, we share our AI models and agents with trusted partners to support progress in three key areas: prevention, detection and response. Here are some examples:

To prevent:

To ensure that our models, like Gemini, are unthreatening and useful to experts, we follow a four-step security process: threat modeling, assessments, mitigation and monitoring.

We work closely with internal biologists, security experts and external partners to understand potential threats, test our models against them, and build strong defenses.

Additionally, we are working to adapt our SynthID watermarking technology to biology, which could assist DNA synthesis providers screen potentially risky biological sequences generated by AI.

Detect:

We assist make pathogen surveillance more cost-effective. For example, our AlphaEvolve agent can optimize the algorithms used to create and analyze metagenomic sequencing data, helping detect modern outbreaks faster. This optimization allows for faster and more exact DNA analysis, making it cheaper to track diseases around the world on a gigantic scale.

We are also exploring how technologies such as AlphaGenome and Protein function annotation can be used to detect and characterize pathogens from sequence data, identifying modern patterns and emerging threats faster than conventional methods.

React:

Building on AlphaFold’s scientific impact, we provide trusted researchers with access to the latest Google DeepMind artificial intelligence systems to accelerate the design of vaccines and other countermeasures to both known and emerging threats.

To support government and non-profit organizations during modern epidemics, Isomorphic Labs has established a specialized unit tasked with rapidly deploying a drug design engine to design medical countermeasures that could address both naturally occurring pandemics and the potential threats arising from the misuse of advanced artificial intelligence. Collaborating with governments and health authorities around the world to develop a variety of diagnostic and therapeutic strategies enables the Isomorphic Labs Drug Design Engine to have a real impact on biological immunity.

This is a intricate, long-term effort that is part of our broader approach to managing potential chemical, biological, radiological and nuclear (CBRN) threats, consistent with the proactive mitigation measures and stringent assessment protocols included in our Border Security Framework.

We are committed to being open and collaborative – with biosecurity labs, governments and the broader scientific community – to ensure AI is developed and deployed responsibly to assist protect against one of the greatest threats to society.

To learn more about our work and our call for modern partnerships, see this article full update on our approach.

Using Gemini as your primary intelligence

The Gemini model powering ATL Saathi provides the foundational intelligence to transform tinkering labs into AI-powered discovery environments. The ability to create concise instructional materials including AI infographics, video overviews, and interactive quizzes for core curriculum modules helps educators easily navigate sophisticated training materials and seamlessly absorb micro-learning content.

Our latest Gemini 3.5 Flash model instantly generates classroom-relevant, curriculum-aligned project ideas to pique student curiosity and provides teachers with detailed assembly instructions, wiring diagrams and safety precautions for students’ unique problem statements.

Looking to the future

We are implementing ATL Saathi in an initial cohort of 100 pilot schools across the country. We hope that by adopting AI-powered tools and microlearning formats, educators will see significant reductions in administrative burden, improved productivity, and increased readiness and comfort levels to support students tinker and innovate.

By shifting the burden of administrative costs and curriculum translation to AI, we free our teachers to do what they do best: mentor, inspire and lead. Let’s tinker together and build the future.

Today, Google DeepMind and A24 announce a first-of-its-kind research-focused partnership. This collaboration brings together a world-leading research laboratory with the industry’s most leading film studio to support artists develop fresh workflows and techniques. Thanks to this, the tools of the future will be shaped by the creators who apply them.

This partnership creates a deep R&D collaboration between A24 and Google DeepMind, spanning multiple projects over time. By anchoring Google DeepMind innovations directly in the original process, A24 and its creators can support shape fresh technologies aligned with their vision and expand storytelling possibilities. This hands-on collaboration provides Google DeepMind with invaluable feedback and guidance from leading artists. Additionally, Google invested in A24.

Looking to the future, the partnership marks the beginning of a shared journey rooted in research and shared curiosity. While the initial focus is on bridging the gap between cutting-edge technology and next-generation entertainment, the specific goals, technical deliverables and original milestones of this initiative will evolve over time. As researchers at A24 and Google DeepMind collaborate to test, iterate, and build, this partnership aims to expand the possibilities for the future of entertainment.

Limitations:

• Omni currently offers 10-second video generations, with longer ones coming soon.
• Uploading audio references and scene extensions is not yet supported in the Gemini API for this model.
• Video references up to 3 seconds long are accepted by the API schema, but are currently not processed correctly by the model.
• Character consistency during scene changes or panning movements has some limitations, but we are working to improve this.

Gemini Omni is available in public preview starting today on Google AI Studio and Gemini API. For a complete list of model capabilities and region-specific restrictions, please check the developer documents.

Build both models today

The real magic happens when you put these models together. Apply Nano Banana 2 Lite as a model to quickly generate an image, then pass that image as a reference to Gemini Omni Flash to animate it into high-quality video. Additionally using Interaction API For these multi-turn experiences, you can maintain session history and context so that users can chain together up to three consecutive edits.

To support you get started, we’ve created some demo apps that you can remix and let you see how to combine Nano Banana 2 Lite and Gemini Omni Flash into one process.

How do we protect internal systems against increasingly capable and imperfectly adapted artificial intelligence

AI agents are changing our relationship with technology. Autonomously performing sophisticated tasks – from cyber defense to scientific discovery and product development – these systems open a novel era of productivity. In the US alone, AI agents could earn $2.90 trillions in economic value by 2030.

As agents become more capable, they also require more sophisticated security. That’s why we developed ours AI control roadmap: a framework for building and managing the advanced AI we deploy at Google. This “defense in depth” approach, which could serve as a model for the broader industry, goes beyond classic model matching by adding a key layer of system-level security that provides confidence even when alignment is imperfect.

Understanding AI control

Our approach to security starts with a forceful foundation, including classic security features such as sandboxing, endpoint security, and rapid injection resistance. Moreover, the AI ​​control roadmap uses model adaptation, i.e. training the AI ​​to be inherently secure and helpful, as its main defense. It provides an additional layer of security by treating internal agents as potentially misaligned, providing confidence even if the match is imperfect.

Think of it like a dual-control driving instructor. The instructor trusts the student but remains ready to take over the wheel or hit the brake if a mistake occurs. Similarly, our AI control system grants permissions to AI agents based on their verified behavior, allowing us to build trust through controlled, incremental access.

Our roadmap is based on a elementary but cautious assumption: What happens if a highly capable AI agent is inconsistent with our goals and begins to act in unexpected ways? To address this, our plan focuses on three key areas.

First, guided by cybersecurity best practices, we have developed a novel threat modeling framework for artificial intelligence. The action plan treats untrusted AI agents as potential “insider threats” – much like a company would treat a potential rogue employee who already has access to the office. Based on industry standards AT&CK DITS framework, a globally available knowledge base about enemy tactics and techniques, based on real-world observations, we divide potential attacks into smaller tactics and techniques. This allows us to methodically track risk, detect bad behavior before it causes damage, and safely conduct drills to see if our defenses actually hold up.

Second, we are developing a plan to implement AI control mitigation measures to address these threats.

• Problem identification (detection): Monitoring is the basis of AI control. We employ other trusted AI systems as “supervisors” that constantly check an employee’s reasoning, actions and plans to ensure they are on track.
• Problem management (prevention and response): If a supervisor sees an agent trying to do something harmful, they can step in to block the action before damage occurs.

Finally, to ensure the functionality of our systems, we continually measure performance based on three critical metrics: reach (the portion of traffic monitored), recall (the portion of abnormal behavior detected), and response time.

Scaling security as AI gets smarter

As AI models continue to evolve, our defenses must also become stronger. In our AI roadmap, we map security protocols to measurable milestones in AI capabilities on two key fronts:

• Ability to avoid detection: Today we monitor the model’s observable train of thought. However, as models learn to hide their reasoning—either by (knowing they are being monitored) or (reasoning without using observable text)—merely reading their verbalized reasoning is not enough. We will need to analyze their behavior and potentially check the inner workings of the model.
• Capacity for potential harm: For low-risk and reversible activities, we can rely on reviewing transcripts with a delay and correcting problems quickly after they occur (asynchronous response). For particularly risky activities, such as major cyber attacks, we need to move to real-time prevention, blocking malicious activities before they can be carried out (synchronous response). In other words, our response strategy must scale based on the severity of the threat.

Secure operate of your computer in Flash version 3.5

To reduce the risk of rapid injection of agents operating in live environments, we operate targeted adversarial computer skills training in Gemini 3.5 Flash. We also provide two optional enterprise security systems that enable enterprises to:

• Require explicit user confirmation for sensitive or irreversible actions.
• Automatically stop tasks when indirect prompt injection is detected.

Using a “defense in depth” approach, we encourage developers to combine these features with a secure sandbox, human-in-the-loop verification, and tight access controls. Additional information on security measures can be found in our best practices documentation.

We are already seeing that customers are adding value by using a computer. Here’s what some of them have to say:

The UK government’s up-to-date AI planning prototype built using Gemini aims to halve the time it takes to process homeowner applications

Governments around the world are exploring how artificial intelligence can deliver better public services faster. The UK is working to build 1.5 million up-to-date homes by 2029, but work is often slowed down by local planning authorities due to high levels of paperwork and backlogs. To aid build Britain, we are working with the UK Government to dramatically reduce processing times for household planning applications. Our aim is to aid officials reduce the time it takes to decide on applications by 50%, freeing up planners’ time to build more homes. We’re excited to see how our national AI partnerships, which aim to support a rethink of public services to create more resilient societies, can aid the UK grow faster.

Co-creating a up-to-date AI planning tool enabling faster decision-making

After successful launch Extract – a tool created with Gemini by the UK Government’s AI Incubator (i.AI) – to aid local governments transform elderly planning documents into clear, digital data – we are working together to develop a up-to-date, sophisticated planning prototype for local governments.

Google DeepMind is working with the UK Government, Google Cloud, the department and local planning authorities in Barnet, Dorset and Camden to co-develop an AI-powered prototype that acts as a highly skilled assistant to planning officers tackling challenging data extraction and case analysis operations. Close cooperation will enable us to build an optimized tool, tailored to the unique needs and everyday challenges that planners face on a daily basis. Following early trials in Barnet, Camden and Dorset, the Government plans to make the up-to-date AI planning tool available to all councils across the country from 2027.

In a typical up-to-date planning application, officers spend hours comparing policy documents, historical files and PDFs. This manual process creates a grave bottleneck. This is particularly challenging given that household applications make up almost 70% of planning applications each year. By reducing time spent on plain matters such as loft conversions or extensions, the prototype being tested in Barnet, Camden and Dorset could aid planning professionals focus more on complicated public benefit applications.

The AI ​​scheduling tool will streamline routine tasks using AI by:

• Data consolidation: Pre-process backlogs, highlight data gaps and extract key site information so planners can view everything on one screen.
• Identifying local policies: Highlighting relevant national and local policies, pre-assess compliance and provide correct citations for review by the official.
• Review summary: Reviewing individual consultation letters to identify key concerns or precedents.
• Preparation of assessments: Creation of the first draft of the final report, containing justification and proposed conditions.

Most importantly, the planning officer has full control as the final decision maker. They review every line generated by the tool, edit the justification, and retain the authority to approve or reject the request. To ensure accountability, the prototype records its work at every step, creating a clear chain of thought and a solid audit trail for every decision to support the planning professional.

The results of this pre-registered study suggest that AI can be a powerful pedagogical partner—not replacing teachers, but extending their reach. This study is part of ours constant effort build a global evidence base on the impact of artificial intelligence on teaching and learning.

Beyond the Answer Engine: Protecting Critical Thinking

A common concern is that generative AI could become a shortcut for students, potentially bypassing the demanding but necessary cognitive effort required for deeper learning. Guided learning aims to solve this problem: it is the result of many years of research and work in our company Find out LM’s efforts be pedagogically grounded and specifically tailored to prioritize building understanding before providing direct answers.

Data from Sierra Leone suggests this approach is working. An analysis of over 113,000 interactions exchanged during our trial period found that in 91.4% of conversations, students used the tool to build conceptual understanding rather than simply looking for solutions. Gemini responded with follow-up questions in 76% of its communications, but provided direct solutions in just 2% of cases. This “Socratic” interaction ensures that the student is left with the cognitive bulky lifting.

Teacher-led intervention

The success of this study was based on a partnership between AI and teachers that kept teachers at the center of the experience. Teachers designed lessons, set goals, and led classroom discussions that fueled learning.

In focus groups, teachers said Gemini also supported their own professional development. Using the lesson preparation tool, they discovered novel ways to explain familiar topics such as fractions. Many described a shift from “lecturers” to “facilitators” who moved around the classroom to support pairs of students in their own learning journeys.

To facilitate others implement similar programs, we are sharing the file teacher training guide with materials created in collaboration with Fab AI, including the specific protocols used in this study.

Impact measurement

The quantitative results were significant. Students using guided learning saw a +0.258 raise in standard deviations in their math scores compared to the control group. In practice, this equates to approximately 1.2 to 1.7 years of typical learning progress achieved during the eight-week trial period.

Students in classrooms where their teachers incorporated Gemini into about half of their lessons to meet the 12-hour target during the trial period saw even greater progress—progress of approximately 1.8 to 2.5 years. Engagement was also exceptionally high: 69% of students met or exceeded usage goals, well above the five percent typical for voluntary learning technologies (known as “The five percent problemThis means that students were not only engaged, but also more willing to come to classes.

Beyond the numbers, we also saw a profound change in behavior. Students reported enjoying mathematics more and actively engaging in learning beyond regular instruction. Most importantly, over time, their conversations and questions became more learning-oriented and moved towards developing skills rather than looking for immediate solutions. In particular, the number of skill-building questions increased to 90% in the last week – from 68% in the first week – while the number of solution-finding questions dropped from 25% to 10%, proving that students not only wanted the answer, but also understood how they got there.

The path forward

While these results are promising, they also highlight the challenge of the “achievement gap.” Although most students benefited, the greatest gains were among those who entered the study with better math skills. This highlights an important need: to offer the tools that provide the greatest benefit to the students who need them most.

Looking ahead, we plan to expand this research to other countries and delve deeper into areas such as metacognition and relational intelligence to capture a more holistic view that explores the nuanced complexities of learning. By combining the relational foundations of the teacher-led student classroom with the personalized, empowering capabilities of artificial intelligence, we can help ensure that technology is a bridge to meaningful learning opportunities for all.

¹ We also received support from Google.org and the Gates Foundation to conduct the study. EducAid, laterite AND Oxford MeasurEd he also cooperated with us.