Gemini 3: Introducing the latest Gemini AI model from Google

Today we’re releasing Gemini 3 – our most intelligent model that helps you bring any idea to life.

blog.google

Gemini 3: Google’s Revolutionary AI Model – Comprehensive Analysis

Executive Overview

Announced on November 17, 2025, Gemini 3 represents Google’s most significant advancement in artificial intelligence since the Gemini series inception in December 2023. This third-generation model establishes transformational benchmarks in multimodal reasoning, agentic capabilities, and enterprise-grade code generation. The flagship Gemini 3 Pro model integrates text, images, video, audio, and code processing into a unified system demonstrating unprecedented depth and contextual awareness across modalities.

The technological achievement is quantifiable and substantial. Gemini 3 Pro achieves a historic 1501 Elo score on the LMArena Leaderboard, marking the first AI system to cross the 1500 threshold. This breakthrough signals fundamental advancement rather than incremental progress in how AI systems approach complex reasoning tasks. The model demonstrates genuine PhD-level cognitive capabilities, scoring 37.5 percent on Humanity’s Last Exam without tool assistance and achieving 91.9 percent on GPQA Diamond, a graduate-level knowledge assessment.

Complementing the base Pro variant, Gemini 3 Deep Think mode extends reasoning capabilities even further through extended inference-time deliberation. This enhanced mode reaches 41 percent on Humanity’s Last Exam and 45.1 percent on ARC-AGI-2, a benchmark specifically designed to test novel problem-solving abilities that cannot be solved through pattern matching from training data. These scores demonstrate genuine reasoning capabilities rather than memorization, addressing one of the most challenging frontiers in contemporary AI development.

Technical Architecture and Innovation

Gemini 3 employs a sophisticated Sparse Mixture of Experts architecture enabling efficient specialization and computational scale. This approach activates different parameter subsets depending on input characteristics, allowing specialized expertise for text, images, video, audio, and code without requiring every parameter to process every token. The architectural improvements include enhanced MoE backbone tuning and expert routing optimizations not present in previous generations.

A unified multimodal stack better fuses cross-modal reasoning compared to Gemini 2.x models. Rather than treating different modalities as separate processing streams that merge at output, Gemini 3 integrates multimodal understanding throughout the entire reasoning process. This architectural decision proves particularly important for benchmarks like ARC-AGI-2, where visual patterns must inform logical reasoning in tightly coupled ways.

Deep Think mode represents a revolutionary operational capability, extending internal deliberation through extended hypothesis testing and alternative evaluation before responding. This inference-time reasoning extension operates separately from base model training, allowing users to choose between fast standard responses and slower, more deliberative outputs depending on specific task requirements. Users report response times of 10-15 seconds for complex reasoning tasks, with dramatic accuracy improvements justifying the latency trade-off for appropriate applications.

Streamed token processing enables handling extremely long documents without loading entire files simultaneously. Gemini 3 reads progressively while retaining relevant parts of conversations, maintaining sustained reasoning over millions of characters without coherence degradation. This design proves especially effective for multi-file uploads and dynamic Workspace projects, such as simultaneously analyzing corporate financials in Sheets while referencing related documentation in Docs.

The model supports a context window of up to one million tokens, leading the industry on long-context benchmarks. This capacity enables consuming entire codebases, lengthy research papers, extended video content, and multi-document collections without losing coherence. Gemini 3 generates up to 128 output tokens per second, ensuring responsive user experiences even during complex reasoning tasks.

Performance Benchmarks and Quantified Achievements

Gemini 3’s performance across standardized evaluations provides compelling evidence of advancement over predecessors and competitors. On Humanity’s Last Exam, designed to test PhD-level reasoning, Gemini 3 Pro’s 37.5 percent score represents an 11 percentage point improvement over GPT-5.1, which achieves 26.5 percent. Claude 4.5 Sonnet scores merely 13.7 percent on the same assessment, placing Gemini 3 substantially ahead of both primary competitors in abstract reasoning tasks.

For visual reasoning and novel problem-solving, ARC-AGI-2 results demonstrate unprecedented capability. Gemini 3 Pro reaches 31.1 percent, representing a massive leap from Gemini 2.5 Pro’s 4.9 percent and well ahead of Claude Sonnet 4.5. The benchmark specifically tests model ability to solve novel, unfamiliar challenges through genuine reasoning rather than pattern matching from training data. Gemini 3 Deep Think mode extends this to 45.1 percent, demonstrating unparalleled capability in tackling problems requiring extended contemplation.

Multimodal excellence manifests across diverse benchmarks. Gemini 3 achieves 81 percent on MMMU-Pro and 87.6 percent on Video-MMMU, establishing new standards for visual reasoning and video comprehension. On SimpleQA Verified, which measures factual accuracy, Gemini 3 reaches 72.1 percent, demonstrating substantial progress in reliability—a critical requirement for enterprise deployment.

Mathematical reasoning shows dramatic improvements. Gemini 3 Pro sets a frontier with 23.4 percent accuracy on MathArena Apex, representing more than a twenty-fold improvement over its predecessor on this exceptionally difficult benchmark. Most competing models score below five percent on competition-level mathematical challenges, making Gemini 3’s performance a genuine breakthrough rather than incremental advancement.

In coding and software engineering domains, Gemini 3 tops the WebDev Arena leaderboard with 1487 Elo and scores 76.2 percent on SWE-bench Verified, evaluating coding agents on actual GitHub issues. This represents a 16.6 percentage point improvement over Gemini 2.5 Pro, translating directly into fewer manual fixes and faster development cycles for professional developers.

Long-horizon planning capabilities demonstrate dramatic improvements on Vending-Bench 2, where Gemini 3 Pro achieves a mean net worth of $5,478, substantially outperforming GPT-5.1’s $1,473. This benchmark simulates year-long business strategy and resource management, testing whether models maintain consistent decision-making without drifting off task—a critical capability for production autonomous systems.

Security Certifications and Trust Framework

Gemini 3 has undergone the most comprehensive safety evaluations of any Google AI model to date, establishing new internal standards for pre-release validation. The platform achieved ISO 27001, ISO 27017, ISO 27018, and ISO 27701 certifications, covering information security management, cloud security, personal information protection in cloud environments, and privacy information management.

SOC 2 Type II and SOC 3 compliance certifications provide independent validation of security controls, availability, processing integrity, confidentiality, and privacy. These assessments, conducted by accredited third-party auditors, confirm that Google’s controls operate effectively over extended observation periods rather than point-in-time snapshots.

ISO 42001 certification represents a significant milestone: Gemini is the first generative AI offering for productivity and collaboration to achieve this world’s first international standard for Artificial Intelligence Management Systems. The certification validates that Gemini has been developed, deployed, and maintained responsibly with appropriate ethical considerations, data governance, and transparency measures.

HIPAA compliance enables healthcare organizations to use Gemini with protected health information subject to appropriate business associate agreements. FedRAMP High authorization package submission occurred in late 2024, positioning Gemini for deployment within U.S. federal government agencies with high-impact security requirements.

Google’s Frontier Safety Framework guided Gemini 3 development from conception through release. This structured risk assessment process identified severe risks, modeled potential harms, conducted assessments across critical domains, and implemented mitigations before deployment. The framework addresses CBRN risks, cybersecurity threats, autonomous replication, manipulation capabilities, and AI research sabotage potential.

External safety testing for CBRN risks found that Gemini 3 Pro offers minimal uplift to low-to-medium resource threat actors compared to established web baselines. Potential benefits are largely restricted to time savings for technically trained users, with minimal utility for less technically trained users due to lack of sufficient detail compared to open sources.

Adoption Statistics and Market Penetration

Within the first 72 hours of launch, Gemini 3 achieved widespread integration across Google’s ecosystem. The Gemini app, which had already reached 650 million monthly users by November 2025, immediately received Gemini 3 access for all users. AI Overviews in Google Search, serving two billion users monthly, now leverage Gemini 3 for more complex reasoning and dynamic generative UI experiences.

Enterprise adoption indicators demonstrate rapid uptake. Over 70 percent of Google Cloud customers use AI services as of late 2025, with Gemini 3 availability through Vertex AI positioning it for immediate deployment across this established base. Development communities responded quickly, with 13 million developers having built applications using Google’s generative models providing a ready ecosystem for integration.

Higher education institutions demonstrate strong momentum, with more than 1,000 U.S. colleges and universities having integrated Gemini for Education, reaching over 10 million students. Early adopters report dramatic productivity gains: one enterprise pilot program documented more than 40 percent time savings in active Workspace activities, with projected annual savings of 110,000 hours across a subset of users.

Third-party platform integrations expanded rapidly post-launch. Gemini 3 became available through Cursor, GitHub Copilot, JetBrains IDEs, Replit, and other development environments within days of announcement. GitHub reported 35 percent higher accuracy in resolving software engineering challenges with Gemini 3 Pro compared to Gemini 2.5 Pro in early VS Code testing, accelerating developer interest.

Geographic reach spans over 100 countries across North America, Europe, Asia, Africa, and Latin America, with availability through multiple access tiers including free consumer access, Google AI Pro and Ultra subscriptions, and enterprise licensing through Workspace and Vertex AI.

Enterprise Success Stories and Business Impact

Early production deployments reveal measurable transformations across diverse sectors. Wayfair, the home goods retailer, achieved a 10 percent boost in response relevancy for complex code-generation tasks requiring data retrieval, coupled with 30 percent reduction in tool-calling mistakes. These improvements translate directly to customers receiving correct answers more frequently with reduced latency.

WRTN, a Korean AI services company, leverages Gemini 3 across their operational spectrum, from Story Generation to Companion Chat, Memory Management, and B2B Agent Projects. The company highlights Gemini 3’s multilingual stability, particularly in high-fidelity languages like Korean, where each model iteration becomes dramatically more natural and stable across all domains—critical for agentic planning workflows.

Figma integrated Gemini 3 into their design platform, with Chief Design Officer noting that the model’s state-of-the-art reasoning and multimodal understanding enable new possibilities for translating design intent into executable code. Enhanced zero-shot generation capabilities allow development teams to rapidly generate well-organized wireframes and high-fidelity frontend prototypes with superior aesthetics.

JetBrains challenged Gemini 3 Pro with demanding frontline tasks from generating thousands of lines of frontend code to simulating operating-system interfaces from single prompts. The model demonstrated more than 50 percent improvement over Gemini 2.5 Pro in solved benchmark tasks, leading to integration into JetBrains’ products delivering smarter, more context-aware experiences to millions of developers.

Box, the enterprise content management platform, implemented Gemini 3 for document understanding and workflow automation. Rakuten served as early alpha tester, validating enterprise readiness through production testing in real-world environments. These implementations demonstrate practical reliability beyond controlled benchmarks.

Academic institutions report transformative outcomes. John Jay College collaborated with Google.org on a predictive AI model identifying students at risk of dropping out. Using 75 indicators including attendance patterns and grade variations, the AI creates risk scores enabling one-on-one coaching before students face trouble. Senior graduation rates rose from 54 percent to 86 percent in three years—an increase nearly unprecedented in higher education.

Arizona State University achieved four times more accurate enrollment predictions, boosting online registrations by 52 percent. University of Maryland graduate finance students built credit risk analysis tools processing data from 40 banks and 17 fintech companies, creating systems rating credit risk management effectiveness across financial institutions—work that previously required expensive Bloomberg terminals and specialized software.

Google Antigravity: Agentic Development Platform

Google Antigravity demonstrates Gemini 3’s multi-agent coordination capabilities in production environments. The platform introduces distinct modes: Editor view for hands-on coding with an agent sidebar, and Manager view as mission control for orchestrating multiple agents across workspaces simultaneously.

Manager view enables asynchronous agent operations. Developers can dispatch five different agents to work on five separate bugs simultaneously, effectively multiplying throughput without requiring sequential attention to each task. Each agent operates autonomously while generating artifacts—task lists, implementation plans, code diffs, screenshots, and browser recordings—providing verifiable evidence of progress.

Agents in Antigravity possess direct access to editor, terminal, and browser surfaces. This architectural decision transforms AI assistance from a tool in the developer’s toolkit into an active partner capable of autonomous planning and execution. Agents can edit files, run commands, validate code through browser testing, and iterate based on test results without continuous human supervision.

The artifact system addresses critical challenges in multi-agent workflows: trust and transparency. Rather than scrolling through raw tool call logs, developers review tangible deliverables explaining agent reasoning at a glance. When something appears incorrect, developers leave feedback directly on artifacts—similar to commenting on documents—and agents incorporate input without stopping execution flow.

Cross-agent learning represents another coordination innovation. Antigravity treats learning as a core primitive, allowing agents to save useful context and code snippets to a shared knowledge base. Subsequent agents access this accumulated knowledge, improving performance on future tasks through organizational memory persisting beyond individual sessions.

Real-world applications demonstrate practical value. Flight tracker application examples show Gemini 3 independently planning full implementation, coding frontend and backend components, integrating with external APIs like AviationStack, generating test data, validating execution through browser automation, and producing documentation—all while developers review progress through artifact updates rather than micromanaging each step.

Pricing and Accessibility

Google structures Gemini 3 access through multiple tiers accommodating different user types and volume requirements. The free tier provides generous capacity for individual developers, hobbyists, and small-scale applications. Users access Gemini 3 Pro with 5-15 requests per minute, 25-250 requests per day, and 250,000 tokens per minute throughput—sufficient for prototyping and low-traffic projects.

Consumer access through the Gemini app offers free usage with rate limits or subscription options. Google AI Pro subscription at approximately $20 monthly provides increased rate limits and priority access. Google AI Ultra subscription at approximately $30 monthly adds Gemini Agent features, Deep Think mode access, and Workspace integration.

Developer API pricing operates on pay-as-you-go terms. Gemini 3 Pro Preview costs $2.00 per million input tokens and $12.00 per million output tokens for prompts up to 200,000 tokens. Prompts exceeding 200,000 tokens incur $4.00 per million input tokens and $18.00 per million output tokens.

Context caching reduces costs for repeated operations. Storing frequently accessed context costs $1.00 per million tokens per hour—substantially cheaper than reprocessing identical inputs across multiple requests. Applications with stable reference materials benefit significantly from caching.

Enterprise pricing through Gemini for Workspace ranges from $20 per user monthly for Business tier to $30 per user monthly for Enterprise tier. These subscriptions require underlying Google Workspace subscriptions as prerequisites. Enterprise tier includes full Gemini 1.5 Pro access, enhanced security controls, and advanced features.

Vertex AI enterprise deployment follows similar token-based pricing with enhanced enterprise features. Organizations benefit from enhanced security certifications, SLA commitments, dedicated support, and integration with existing Google Cloud infrastructure.

Competitive Positioning

Gemini 3’s competitive position reflects genuine technological leadership. The 1501 Elo LMArena score—the first model to cross 1500—represents transformational advancement. While GPT-5.1 maintains competitive pricing and Claude 4.5 Sonnet demonstrates superior performance on specific coding tasks, Gemini 3’s combination of reasoning depth, multimodal excellence, and ecosystem integration creates unique value unavailable from competitors.

First-day Search integration represents a strategic differentiator. Gemini 3 launched simultaneously across consumer products including AI Mode in Search, marking the first time Google shipped a flagship model to Search on day one. This coordination demonstrates operational maturity competitors struggle to match.

The Google ecosystem depth provides structural advantages. Native integration with Gmail, Docs, Sheets, Drive, Calendar, Meet, Maps, YouTube, and Android creates workflow continuity unavailable from point solutions. Organizations already standardized on Google Workspace gain AI capabilities without switching costs or integration complexity.

Agentic development platform leadership through Google Antigravity establishes a new category. While competitors offer coding assistants, Antigravity’s agent-first architecture with dedicated Manager view for orchestrating multiple autonomous agents represents qualitative innovation. The artifact system for verifiable work products and cross-agent learning through persistent knowledge bases create defensible differentiation.

Educational market dominance through Gemini for Education positions Google uniquely. Serving over 10 million college students across 1,000-plus U.S. institutions with enterprise-grade data protections free of charge creates early-career adoption patterns persisting into professional life.

Multimodal excellence across video, audio, and visual understanding establishes technical leadership. The 87.6 percent Video-MMMU score substantially exceeds competitors, enabling use cases like sports coaching analysis, educational video processing, and surveillance applications where competitors cannot match accuracy.

Long-context leadership with one million token windows maintained across sustained reasoning tasks addresses enterprise requirements for processing entire codebases, comprehensive document collections, and extended conversations without losing coherence. Competitors with 200,000 token limits face architectural constraints requiring workarounds that introduce latency and complexity.

Enterprise Implementation and ROI

Enterprise ROI studies reveal substantial productivity gains. One major U.S. organization analyzed their Gemini pilot across several thousand employees, revealing that certain business units achieved more than 40 percent time savings in active Workspace activities, averaging over 30 minutes saved per week per user.

Projected annual time savings of 110,000 hours translated to approximately 1.8 million dollars in financial return for a single business unit. These gains came from reduced email drafting time, faster document creation, automated meeting summaries, and enhanced data analysis capabilities.

Development efficiency improvements show measurable impact. Organizations deploying Gemini 3 for coding report 35 percent higher accuracy in resolving software engineering challenges compared to previous generations. This translates to fewer manual fixes, reduced debugging time, and faster feature delivery.

Customer support cost reductions emerge from automation of routine inquiries. Organizations report handling 30-50 percent more support volume without proportional staff increases, as Gemini-powered systems resolve straightforward questions while escalating complex issues to human agents.

Content creation acceleration proves valuable for marketing teams. Teams generating blog posts, social media content, and email campaigns report 50-70 percent time reduction compared to manual creation. A content team producing 100 pieces monthly reduces time-per-piece from 4 hours to 1.5 hours, saving 250 hours monthly.

Educational institutions calculate ROI through improved student outcomes. John Jay College’s graduation rate improvement from 54 percent to 86 percent over three years translates to hundreds of additional students completing degrees annually, representing millions of dollars in tuition revenue retention and enhanced lifetime earnings.

Future Roadmap and Strategic Vision

Google plans releasing additional Gemini 3 series models, enabling broader AI capabilities across expanded use cases. While specific variants and timelines remain unannounced, the roadmap likely includes specialized versions optimized for particular domains, similar to Gemini 2.0 Flash variants.

Gemini 3 Deep Think mode undergoes extended safety evaluations before public release to Google AI Ultra subscribers in coming weeks. This phased rollout reflects commitment to responsible deployment, ensuring the most powerful reasoning mode receives thorough validation.

On-device AI expansion through Gemini Nano 3 will bring offline capabilities to Android devices. Integration of advanced AI features directly into mobile operating systems enables privacy-sensitive applications processing data locally rather than transmitting to cloud servers.

Autonomous task completion capabilities through Gemini Agent will expand beyond current Gmail organization and local services booking. The system demonstrated long-horizon planning capabilities, suggesting potential for complex business process automation spanning multiple applications.

Gemini Enterprise platform evolution will enhance agent discovery, creation, sharing, and execution capabilities. Organizations increasingly need centralized governance for agent proliferation, and Google’s roadmap emphasizes secure, compliant environments for enterprise agentic deployments.

Integration depth across Google products will continue expanding. Current availability spans Search, Gemini app, Workspace applications, Maps, and YouTube, but substantial opportunity remains for deeper embedding into Android, Chrome, Cloud Platform services, and specialized products.

Strengths and Competitive Advantages

Gemini 3’s reasoning depth represents its most significant strength. The 1501 Elo score on LMArena—the first model to cross 1500—provides quantifiable evidence of advancement beyond incremental improvement. Performance on Humanity’s Last Exam demonstrates genuine PhD-level reasoning capability.

Multimodal understanding sets new industry standards. Achieving 87.6 percent on Video-MMMU establishes Gemini 3 as the strongest model for video comprehension among publicly available systems.

Real agentic execution through Antigravity and Gemini Agent moves beyond chatbot interactions to task-level automation. The system’s ability to autonomously plan, execute, and verify complex multi-step workflows demonstrates practical reliability for production deployment.

Integration with Google’s ecosystem provides unique advantages. Native support for Gmail, Docs, Sheets, Drive, Calendar, and Meet enables seamless workflow automation across productivity tools used by billions.

Developer tooling represents another differentiator. Google Antigravity’s agent-first IDE architecture and comprehensive artifact system for verifiable work products provide developers flexibility unavailable in single-vendor toolchains.

Limitations and Considerations

Response latency in Deep Think mode presents trade-offs. Users report 10-15 second response times—acceptable for thoughtful analysis but frustrating for rapid-fire interactions. Mitigation involves using standard mode for routine queries and reserving Deep Think for situations justifying extended computation.

Conservative safety guardrails occasionally impede creative workflows. The model refuses image generations featuring public figures and exercises caution on ambiguous historical questions, requiring users to reword prompts before receiving responses.

Data freshness limitations affect time-sensitive queries. Training data cutoffs mean the model cannot directly answer questions about events after late 2024 without grounding. The 1,500 free daily grounded requests provide substantial capacity for typical applications.

Resource intensity translates to higher costs for some workloads. Token prices are slightly elevated compared to lighter models. For budget-constrained applications, using Gemini 2.0 Flash for routine tasks while reserving Gemini 3 for complex reasoning optimizes cost-performance trade-offs.

Niche domain limitations appear in highly specialized jargon. The model occasionally struggles with super-specific terminology in fields like quantum computing or traditional herbal medicine. Subject matter experts must verify outputs in specialized domains.

Offline capabilities do not exist currently. Internet connectivity requirements prevent usage in disconnected environments. Organizations requiring air-gapped deployments must wait for potential future releases supporting local execution.

Market Transformation and Industry Implications

The AI industry is experiencing a fundamental shift from conversational assistants to agentic systems capable of extended autonomous operation. Gemini 3’s emphasis on tool use, long-horizon planning, and multi-step task execution positions Google advantageously for this transition.

Multimodal AI maturation enables new application categories. As models achieve human-competitive performance on video understanding, spatial reasoning, and cross-modal synthesis, use cases previously impossible become practical.

Enterprise AI adoption accelerates as security certifications, compliance frameworks, and governance tooling mature. Organizations previously hesitant to deploy AI for sensitive workloads can now satisfy regulatory requirements through platforms with appropriate certifications.

The total cost of ownership conversation shifts from technology costs to business impact. As API pricing commoditizes and free tiers provide generous capacity, competitive differentiation emerges from implementation effectiveness.

Developer productivity tools evolving toward agent-first architectures require rethinking development practices. Google Antigravity’s Manager view for orchestrating multiple autonomous agents represents future development paradigms.

Education integration at scale creates generational adoption patterns. With over 10 million college students using Gemini for Education, workforce entrants arrive with AI-native work habits and expectations.

Conclusion

Gemini 3 represents Google’s most ambitious and successful AI release to date, establishing new benchmarks across reasoning, multimodal understanding, and agentic capabilities while demonstrating operational maturity to ship flagship models across products on launch day. The 1501 Elo LMArena score, 45.1 percent ARC-AGI-2 performance with Deep Think mode, and 87.6 percent Video-MMMU achievement signal genuine advancement rather than incremental improvement.

The model’s practical impact extends beyond impressive benchmarks to measurable business outcomes: enterprises achieving 40 percent time savings in active Workspace activities, development teams experiencing 35 percent higher accuracy resolving software engineering challenges, and educational institutions driving graduation rates from 54 percent to 86 percent through AI-powered student support. These results validate AI’s transition from experimental technology to mission-critical infrastructure.

Google’s strategic advantages in distribution, ecosystem integration, and infrastructure position Gemini 3 uniquely among competitors. With two billion monthly users accessing AI Overviews in Search, 650 million using the Gemini app, and over 70 percent of Cloud customers deploying AI services, Google possesses unparalleled channels for delivering advanced capabilities at scale.

Challenges remain, including Deep Think mode latency, conservative safety guardrails, slightly elevated pricing, and data freshness limitations, requiring mitigation through thoughtful implementation choices. Organizations must balance reasoning requirements against response time needs, configure grounding for time-sensitive applications, and optimize model selection across workloads to control costs effectively.

For enterprises evaluating AI platforms, Gemini 3 merits serious consideration, particularly for organizations already invested in Google Cloud infrastructure or Workspace applications. The combination of technical excellence, comprehensive security certifications including ISO 42001, favorable total cost of ownership economics, and extensive ecosystem integration creates compelling value propositions.

Educational institutions and individual developers benefit from unusually generous free tier access and dedicated educational programs. These accessibility commitments demonstrate Google’s long-term strategy of building adoption patterns and institutional knowledge that compound as students transition to professional contexts.

The broader trajectory suggests AI’s ongoing transformation from narrow tools to general-purpose reasoning partners capable of understanding context, maintaining long-term plans, coordinating multiple simultaneous workflows, and generating interactive experiences, visualizations, and working software. Gemini 3 advances this vision substantially while remaining transparent about limitations and maintaining strong safety governance through comprehensive evaluations and responsible deployment practices.

Gemini 3: Introducing the latest Gemini AI model from Google

Today we’re releasing Gemini 3 – our most intelligent model that helps you bring any idea to life.

blog.google