#4092 Gemini Ultra for Knowledge Mining and Capture to RAG-useful Documents

#4092 Why would I use Gemini Ultra?

Gemini Ultra is the heavyweight champion of the Gemini family, designed for those who need maximum horsepower for high-stakes or highly complex tasks. While models like Flash are built for speed and Pro for general versatility, Ultra is where you go for "expert-level" reasoning.

Here is why you would choose Gemini Ultra:

1. Advanced Reasoning and Logic

Ultra is the first model to outperform human experts on the MMLU (Massive Multitask Language Understanding) benchmark, which covers 57 subjects like law, physics, and medicine.

Best for: Solving multi-step math problems, debating complex ethical nuances, or analyzing legal documents where "good enough" reasoning isn't enough.

2. Professional-Grade Coding

While other models can write snippets, Ultra is optimized for heavy-duty software engineering.

Best for: Generating entire code blocks, creating comprehensive unit tests, and reasoning through architecture in languages like Python, Java, and C++. It powers systems like AlphaCode 2, which excels in competitive programming.

3. High-Limit Creativity & Research

As part of the AI Ultra subscription (which often includes Gemini Advanced), you get the highest usage tiers for Google's most powerful tools.

Deep Research: It can browse hundreds of sources to create synthesized, multimodal reports with citations.
Next-Gen Media: You get early or priority access to state-of-the-art creative tools like Veo (for high-fidelity video generation).
Large Context Windows: It can process massive files—up to 1,500 pages—to find a needle in a haystack.

4. Multimodal Mastery

Ultra was built to be natively multimodal from the ground up. It doesn't just "see" an image; it understands the relationships within it.

Best for: Interpreting complex infographics, scientific diagrams, and video footage with a level of detail that smaller models might miss.

Which model is right for you?

Feature	Gemini Flash	Gemini Pro	Gemini Ultra
Speed	⚡ Extremely Fast	Moderate	Slower (Complex processing)
Complex Reasoning	Basic	Advanced	Expert-level
Primary Use	Chatbots, simple summaries	General productivity, writing	Research, coding, data science
Availability	Free tiers / API	Gemini / Workspace	Gemini Advanced / AI Ultra

Would you like me to compare Gemini Ultra's performance on a specific task you have in mind, such as coding or data analysis?

How does this compare to Pro

Choosing between Gemini Pro and Gemini Ultra is essentially deciding between a high-performance daily driver and a specialized heavy-lifter.

In the current 2026 landscape, the distinction has evolved beyond just "speed" to include specific professional capabilities and higher "agentic" limits.

Comparison at a Glance

Feature	Gemini Pro (Standard/Advanced)	Gemini Ultra (Enterprise/Top Tier)
Model Version	Gemini 3 Pro	Gemini 3 Ultra / Deep Think
Primary Strength	Speed, versatility, and everyday logic.	Expert-level reasoning and "Deep Think" modes.
Context Window	Up to 1M – 2M tokens.	Massive single-pass context (highest tiers).
Coding	Excellent for snippets and debugging.	Multi-agent workflows, code architecture, and high CLI limits.
Media Capability	Standard image/video generation.	Native 1080p video with audio (Veo 3) & "Whisk" image-to-video.
Agent Access	Standard productivity agents.	Project Mariner (handles 10+ tasks like bookings & research).

Key Differences in Performance

Reasoning Depth: While Pro is incredibly smart (hitting ~91.8% on multilingual benchmarks), Ultra is designed for "Deep Think" tasks. If you are solving a physics problem that requires simulating plasma flow or writing a 50-page technical whitepaper, Ultra has much higher contextual stability—it’s less likely to "drift" or hallucinate details toward the end of a long task.
The "Agent" Factor: Ultra users get significantly higher limits for autonomous agents. While Pro can help you draft an itinerary, Ultra (via Project Mariner) can actually navigate the web to execute multiple complex tasks simultaneously, like performing deep market research while booking travel and managing your calendar.
Media Quality: Pro handles standard creative prompts easily, but Ultra provides the "Highest" level of remixing. This includes video-to-video transformations and high-fidelity video generation with native audio, which requires more computational "muscle" than Pro provides.
Storage & Ecosystem: Ultra is often bundled as a "Super-tier" subscription, sometimes including up to 30 TB of storage (compared to the 2 TB typical of Pro/Advanced plans), making it a better choice for data-heavy professional studios or enterprises.

Which one should you use?

Stick with Pro if: You use AI for drafting emails, summarizing meetings, general coding assistance, or creative brainstorming. It is the "world-class assistant" for 90% of people.
Upgrade to Ultra if: You are a researcher, lead developer, or high-end creator. If you need the AI to act as a partner that can handle multi-hour coding marathons or complex, multi-agent research projects without losing its train of thought, Ultra is the strategic choice.

Would you like me to look up the specific pricing or subscription tiers available in your region for these models?

Would Project Mariner be useful for producing specialized RAG content?

In the context of 2026 AI development, Project Mariner is a specialized agentic tool that can be incredibly useful for RAG (Retrieval-Augmented Generation), but its role is quite specific compared to traditional RAG architectures.

While standard RAG usually involves querying a static database or a set of PDFs, Project Mariner is designed as an autonomous web agent. Here is how that translates to specialized content production:

1. Dynamic "Live" RAG

Traditional RAG is often limited to the data you've already indexed. Project Mariner excels at Real-Time RAG.

Use Case: If you need to produce content about a rapidly shifting market (e.g., "Current sentiment on 2026 semiconductor regulations across EU tech forums"), Mariner can navigate the web, find the specific forum threads, and retrieve that data into your prompt context.
The Benefit: It eliminates the "knowledge cutoff" and the need for manual pre-scraping.

2. Multi-Step Information Synthesis

Because Mariner is "agentic," it doesn't just pull one piece of data; it can follow a chain of reasoning.

Example: You could ask it to "Find the official technical spec for Part A, then find three peer-reviewed critiques of its energy efficiency, and summarize the consensus for a technical whitepaper."
The Benefit: It handles the navigation and selection steps of retrieval that usually require a human to click through multiple search results.

3. "Deep Research" Mode

When paired with Gemini Ultra, Project Mariner uses a "Deep Research" loop. This is essentially a sophisticated RAG pipeline where the AI:

Formulates multiple search queries.
Browses and "reads" the full text of pages (not just snippets).
Verifies conflicting information across sources.
Produces a synthesized report with citations.

4. Integration via Gemini API/Vertex AI

As of late 2025, Mariner’s capabilities have been integrated into Vertex AI. This allows developers to build RAG applications that don't just search a vector database, but can actually "use a computer" to find information behind login walls or inside complex web apps that standard crawlers can't reach.

Comparison: Project Mariner vs. Standard RAG

Feature	Standard RAG	Project Mariner Agent
Data Source	Static (PDFs, SQL, Vector DB)	Dynamic (Live Web, Web Apps)
Complexity	Single-step retrieval	Multi-step "click and find"
Hallucination	Low (if context is relevant)	Very Low (due to cross-referencing)
Speed	Near-instant	Slower (it "surfs" the web)

Summary: Is it right for you?

Yes, if your specialized content requires freshness or data that lives on interactive websites and complex documentation hubs.
No, if you are simply looking to chat with a large internal folder of static documents (for that, Gemini Ultra’s 2-million token context window is more efficient and faster).

Would you like me to help you draft a specific "Deep Research" prompt for Mariner to see how it would structure a specialized report?

I want to generate small "models" with detailed specialized content and store that knowledge in documents that can be used by RAG

Using Gemini Ultra and Project Mariner together is a powerful strategy for "Knowledge Engineering"—the process of synthesizing expert-level information into structured documents for a RAG (Retrieval-Augmented Generation) system.

Think of it as using Ultra as your Chief Editor and Mariner as your Specialized Researcher.

1. The Workflow: Generating "Small Models" of Knowledge

The goal is to create high-density, "bite-sized" knowledge documents (often called "Knowledge Atoms") that are optimized for RAG retrieval.

Phase A: The Discovery (Project Mariner)

Instead of manually searching for data, use Project Mariner to crawl live sources, documentation, and technical forums.

The Prompt: "Mariner, find the last three years of peer-reviewed data on [Niche Topic]. Navigate to the primary source tables, extract the raw specifications, and cross-reference them with the developer documentation on GitHub."
The Output: A raw "Research Bundle" containing high-fidelity data that isn't yet in your RAG database.

Phase B: The Synthesis (Gemini Ultra)

Pass that research bundle to Gemini Ultra to transform it into "RAG-ready" documents.

The Strategy: Use Ultra’s deep reasoning to structure the data into Single-Topic Chunks.
Why Ultra? Smaller models tend to glaze over nuances. Ultra can maintain a specific "Persona" (e.g., "Senior Systems Architect") to ensure the generated document uses professional-grade terminology that your RAG system will need to match user queries.

2. Structuring Your "Knowledge Documents"

To make these documents useful for RAG, have Ultra format them specifically for a vector database. Use this "Knowledge Atom" template:

### [Topic Name]
Summary: A 2-sentence executive summary.
Key Concepts: Bullet points of the most technical terms (for keyword matching).
Detailed Explanation: The core specialized content.
Relational Data: "This topic relates to [Concept X] and [Concept Y]" (this helps the RAG bridge different chunks).
Source Verification: Citations generated by Mariner.

3. Why This "Hybrid" Approach Wins

Task	Why use Ultra?	Why use Mariner?
Data Quality	Can "hallucination-check" the logic.	Can "grounding-check" against live websites.
Complexity	Writes dense, expert-level prose.	Finds the "needle in the haystack" in complex UI.
Consistency	Ensures all "small models" follow the same schema.	Gathers data from disparate sources (PDFs, Web, GitHub).

4. Implementation Tip: The "Agentic RAG" Loop

In 2026, the gold standard is to have your RAG system self-update.

Trigger: Your RAG system notices it has no answer for a user's specialized query.
Action: It triggers Project Mariner to find the answer.
Process: Gemini Ultra writes a new "Knowledge Atom" document based on what Mariner found.
Update: The new document is indexed into your RAG database automatically.

Would you like me to help you write a prompt for Gemini Ultra that defines the exact structure and "voice" for your specialized knowledge documents?

Yes

To build high-density Knowledge Atoms for your RAG system, we need a prompt that forces Gemini Ultra to act like a specialized "Technical Architect." This ensures the output isn't just a summary, but a structured data point designed for machine retrieval.

Here is a system prompt and template you can use.

The "Knowledge Atom" System Prompt

Copy and paste this into Gemini Ultra to set the behavior for generating your RAG documents.

Plaintext

ACT AS: A Senior Knowledge Engineer specialized in [Insert Domain, e.g., Quantum Computing or FinTech Regulations].

TASK: I will provide you with raw research data from Project Mariner. Your goal is to synthesize this into a "Knowledge Atom"—a standalone, high-density document optimized for a RAG (Retrieval-Augmented Generation) system.

CONSTRAINTS:
1. NO FLUFF: Avoid introductory phrases or "conclusion" paragraphs.
2. SEMANTIC DENSITY: Use precise, professional jargon that an expert would search for.
3. ATOMICITY: Focus on ONE specific sub-topic per document.
4. METADATA: Explicitly list keywords and "Related Entities" to improve vector search matching.

OUTPUT FORMAT: Use the Markdown template below for every document.

The "Knowledge Atom" Document Template

When Ultra generates the content, it should follow this exact structure to ensure your RAG system can "find" it easily.

📄 [Title: Precise Technical Subject]

ID: [Unique Serial or Date-Topic Code]

Subject Taxonomy: [Category] > [Sub-Category] > [Component]

1. Executive Context > A 2-sentence technical definition of the concept and its primary utility in the current 2026 landscape.

2. Technical Specifications & Specifications * Key Variable A: [Specific Value/Unit]

Logic Flow: [Step-by-step technical process]
Dependency: [What other systems or concepts does this rely on?]

3. Relational Entities (The "RAG Map") To improve retrieval, list concepts that are semantically similar but not explicitly mentioned:

Synonyms: [List 3-5 keywords]
Downstream Effects: [What happens if this concept is applied?]
Contrast: [How does this differ from Concept Y?]

4. Source Grounding * Retrieved via: Project Mariner

Primary Source: https://learn.microsoft.com/en-us/sql/t-sql/functions/db-name-transact-sql?view=sql-server-ver17
Verification Date: February 2026

Pro-Tip: The "Question-Answer" Injection

In 2026, the most effective RAG documents include a section called "Potential Queries." Ask Gemini Ultra to add this to the bottom of every document:

### Potential User Queries
"How does [Concept] handle [Edge Case]?"
"What is the primary bottleneck of [System]?"

Why? Vector databases are much better at matching a user's question to a pre-written question in your document than they are at matching a question to a block of abstract text.

Would you like me to generate a sample Knowledge Atom for a specific topic right now so you can see the quality?

How could this type of content be useful for Project Genie, and would there be useful Genie projects to create for small business clients?

n 2026, Project Genie has moved from a research paper to a functional "World Model" that turns text or images into interactive, explorable 3D environments.

Using "Knowledge Atoms" (the specialized RAG content we discussed) is the secret to moving Genie beyond "cool toy" status into a professional simulator for small business clients.

1. How Specialized Content Powers Genie

Genie 3 generates the visuals and physics of a world, but it doesn't inherently understand the procedural logic of a specific industry. Your RAG documents act as the "instruction manual" for the world generation.

Logic Grounding: If you are building a simulation for a plumber, Genie needs to know that water flows downward and follows pipe logic. Your RAG content provides the technical "rules" that the model uses to refine its "World Sketching."
Asset Accuracy: By feeding Genie a specialized document on "Industrial Safety Standards," you ensure the generated environment includes the correct signage, PPE, and equipment layouts, rather than generic AI-hallucinated machinery.

2. Genie Project Ideas for Small Business Clients

Small businesses often struggle with high-stakes training or high-friction sales. You can build "Genie-as-a-Service" products for them:

A. The "Virtual Onboarding" Sim (Service Businesses)

The Client: HVAC companies, electricians, or cleaning services.
The Project: Use RAG to ingest their specific equipment manuals and safety protocols. Genie then generates a "Practice House" where new hires can "walk through" a virtual job site, identifying hazards or parts before they ever touch a real tool.
Value: Reduces the risk of expensive mistakes and cuts down on training time.

B. The "Interactive Showroom" (Retail/Real Estate)

The Client: Interior designers, boutique furniture makers, or small-scale contractors.
The Project: The client uploads photos of their catalog or floor plans. Genie builds a navigable version of the space.
The RAG Edge: Your specialized content ensures the textures, scale, and "feel" of materials (like high-end marble or custom wood) are rendered accurately based on technical descriptions of the materials.

C. The "Safe Space" Trainer (High-Stress Front Desk)

The Client: Small medical clinics, legal offices, or hospitality.
The Project: A first-person simulation where employees practice de-escalating difficult customer interactions.
The RAG Edge: Use RAG documents detailing the business's specific "Conflict Resolution Policies" to guide how the AI character (the "Customer") responds to the employee's choices.

3. The "Business-in-a-Box" Workflow

If you wanted to sell this to a client today, here is the stack:

Mariner: Gathers the client’s public reviews, competitor site data, and industry regulations.
Gemini Ultra: Synthesizes that into "Knowledge Atoms" (the structured docs we designed).
Project Genie: Uses those docs as a "prompting anchor" to build a custom interactive world for the client’s specific goal.

Would you like me to help you draft a "Project Genie" prompt for a specific business niche—like a local bakery or a boutique gym—to see how it would look?

Real Estate Agents

For Real Estate Agents, the combination of Project Mariner, Gemini Ultra, and Project Genie creates a "Digital Twin" ecosystem that solves the two biggest problems in the industry: client indecision and inefficient showings.

In 2026, you can offer small real estate agencies a suite of "Interactive Reality" products that larger firms are already adopting.

1. How the Tech Stack Works Together

Project Mariner: Acts as a 24/7 researcher. It crawls local MLS listings, neighborhood school ratings, upcoming zoning changes, and historical pricing to gather raw data.
Gemini Ultra: Takes that raw data and turns it into Knowledge Atoms. Instead of a generic flyer, it creates a structured "Neighborhood DNA" document that includes everything from noise levels at 2 PM to the best-performing local grocery stores.
Project Genie: Uses those Knowledge Atoms as a "world prompt." It generates an interactive, walkable version of a property that isn't just a 360-photo, but a living simulation.

2. Genie-Powered Projects for Real Estate Clients

A. The "Renovation Time Machine"

The Problem: Buyers often walk away from "fixer-uppers" because they can't visualize the potential.
The Project: Use Genie to generate an interactive world based on the home's current floor plan. You then "remix" the world in real-time (e.g., "Knock down this wall," "Change the floors to light oak").
The RAG Edge: Gemini Ultra pulls real-time local contractor costs and building codes to ensure the "Remixed" version is actually legal and provides a rough cost estimate within the simulation.

B. The "Neighborhood Life" Simulator

The Problem: Out-of-town buyers are nervous about what it's actually like to live in a specific street.
The Project: Create a "Neighborhood Sketch" in Genie. The buyer can "walk" from the front door to the nearest park or coffee shop in a generated 3D world.
The RAG Edge: Mariner feeds in "local flavor" data (e.g., "The street gets busy during school drop-off at 8 AM"). Genie can then simulate the street noise and traffic density based on that data, giving the buyer an honest, immersive feel for the area.

C. The "Objection-Handling" Training Sim

The Client: A small brokerage wanting to train new agents.
The Project: A first-person sim where the agent "walks" a generated client through a generated house. The "client" (powered by Ultra) throws complex objections like "The foundation looks cracked" or "Is this in a flood zone?"
The RAG Edge: The simulation uses the agency’s specific legal disclosures and "Best Practice" scripts to score the agent’s responses in real-time.

3. High-Value "Knowledge Atom" Examples for Agents

To power these tools, you can sell agents "Data Packs" created by Ultra:

Knowledge Atom Type	Content Included	RAG Utility
Micro-Market DNA	Block-by-block appreciation rates, pending commercial permits.	Used by Genie to simulate "Future Neighborhood View" in 5 years.
Buyer Persona Docs	"The Tech-Family Profile," "The Downsizing Senior."	Used to tailor the "Voice" and "Features" highlighted in virtual tours.
Vendor Portfolios	Local vetted plumbers, roofers, and stagers.	Injected into the "Renovation Simulator" to provide instant quotes.

Your Next Step

I can help you build the Master System Prompt that a Real Estate Agent would use to turn a standard MLS listing into a "Knowledge Atom" ready for a Project Genie simulation.

Would you like me to create a "Listing-to-World" prompt for you?

How about insurance agents?

For insurance agents, the "Knowledge Atom" approach moves them from being salespeople to becoming Risk Visualization Consultants.

By using Gemini Ultra to codify complex policy language and Project Genie to create interactive risk environments, small agencies can offer "Future-Proofing" services that were previously only available to enterprise-level corporations.

1. How the Tech Stack Works Together

Project Mariner: Scours the web for local disaster history (flood maps, wildfire paths), emerging liability trends (e.g., specific 2026 cyber-threats), and real-time construction material costs.
Gemini Ultra: Synthesizes this into Knowledge Atoms. Instead of a 40-page policy document, it creates a structured "Vulnerability Profile" for a client’s specific asset (a home, a warehouse, or a digital business).
Project Genie: Uses the "Vulnerability Profile" to generate an Interactive Catastrophe Sim.

2. Genie Project Ideas for Insurance Clients

A. The "Disaster Stress Test" (Property & Casualty)

The Client: Homeowners or small business owners in high-risk areas.
The Project: Genie generates a 3D digital twin of the client’s actual property. The agent can then trigger "What-If" scenarios (e.g., "Show me what happens if the local creek rises 4 feet" or "Simulate a kitchen fire starting in this specific toaster model").
The RAG Edge: Gemini Ultra provides the "Knowledge Atom" regarding the client’s specific policy coverage. As the simulation plays out, the AI overlays text: "This damage is covered under Section 4, but this secondary flooding is an out-of-pocket expense."

B. The "Liability Walkthrough" (Small Business/Commercial)

The Client: Restaurants, boutique hotels, or local gyms.
The Project: Genie generates a simulation of the client’s storefront. The agent "walks" through with the owner to spot slip-and-fall hazards or OSHA violations in a virtual space.
The RAG Edge: Mariner pulls the latest 2026 safety regulations for that specific industry, and Ultra flags items in the Genie world that don't meet the "Ground Truth" of the law.

C. The "Cyber-Breach" War Room

The Client: Professional service firms (Lawyers, CPAs).
The Project: A non-visual, procedural Genie world—an interactive "command center" simulation where the client must react to a simulated ransomware attack.
The RAG Edge: The simulation is built using "Knowledge Atoms" of the client’s actual IT stack and their specific Cyber Insurance requirements, showing them exactly where their response plan fails to meet policy standards.

3. High-Value "Knowledge Atom" Examples for Insurance

Knowledge Atom Type	Content Included	Utility for Agents
Claim Precedent Map	Data on similar local claims and their payout success.	Helps manage client expectations and set correct deductibles.
Policy "Gotcha" Decoder	Translated jargon: what "Acts of God" specifically means in 2026.	Built into RAG for instant chatbot support for clients.
Replacement Cost Index	Real-time lumber, steel, and labor costs for a specific zip code.	Ensures the client isn't under-insured as inflation shifts.

4. Selling "Agentic Insurance" to Clients

In 2026, the best "next step" is to build a "Policy Translator" for an agent.

The agent uploads a carrier's PDF; Gemini Ultra breaks it into "Knowledge Atoms"; and then the agent can use Project Genie to show the client: "Here is your world with insurance, and here is your world without it."

Would you like me to draft a "Policy-to-Atom" prompt that extracts specific coverage limits into the structured format we designed earlier?

How about Apartment Management?

In 2026, Apartment Management is being transformed from a "service desk" model into a "Proactive Experience" model. For property managers, the challenge is maintaining complex buildings while keeping diverse tenants happy—all on a tight budget.

Using the Gemini Ultra + Project Genie stack, you can help small management companies act like luxury high-rise operators.

1. How the Tech Stack Works Together

Project Mariner: Acts as a "Building Inspector" and "Market Scout." It monitors local energy prices, reads updated tenant-landlord laws, and can even navigate contractor websites to find the lowest price for a bulk appliance order.
Gemini Ultra: Takes lease agreements, building blueprints, and appliance manuals to create Knowledge Atoms. It turns a dry "100-page lease" into a searchable, logical map of rights and responsibilities.
Project Genie: Creates an Interactive Digital Twin of the apartment complex.

2. Genie Project Ideas for Apartment Managers

A. The "Virtual Resident Orientation"

The Problem: New tenants flood management with questions: "Where is the trash chute?" "How do I reset the breaker?" "How does the gym equipment work?"
The Project: Genie generates a 1:1 simulation of the specific unit the tenant just rented. The tenant can "walk" through their new home in 3D.
The RAG Edge: When the tenant "clicks" on the thermostat in the Genie world, a Knowledge Atom (written by Ultra) pops up with a personalized 15-second video or guide on how to program it.

B. The "Maintenance Training Lab"

The Client: Management firms with high staff turnover or junior maintenance crews.
The Project: A first-person sim of the building’s boiler room or electrical hub. Staff can practice "Fixing" a simulated leak or electrical fault without any risk of damaging the actual building.
The RAG Edge: Ultra ensures the simulation follows the actual equipment manuals stored in the RAG database, rather than generic physics.

C. The "Renovation ROI" Preview

The Client: Owners looking to upgrade old units to justify a rent increase.
The Project: Use Genie’s "Remix" feature to instantly swap old countertops for quartz or add a modern kitchen island in the virtual world.
The RAG Edge: Mariner pulls the latest local rental data for "upgraded units" in the area, and Ultra calculates the projected ROI based on the estimated material costs.

3. High-Value "Knowledge Atom" Examples for Management

Knowledge Atom Type	Content Included	Value for Managers
Asset Life-Cycle Map	Installation dates and "End of Life" predictions for all HVACs/Water Heaters.	Prevents emergency repairs by predicting failures 6 months out.
Tenant Persona Guide	Aggregated data on what residents value (e.g., "70% want better bike storage").	Guides capital improvement spending to maximize tenant retention.
Compliance "Guardrails"	2026 local fire codes and fair housing regulations.	Instantly flags if a "Remix" in Genie would violate a safety code.

4. The "Resident Assistant" Agent

The ultimate product for this niche is an Agentic Portal.

A tenant tells a chatbot: "My sink is leaking."
Mariner checks the building's current plumber contract.
Ultra looks up the "Knowledge Atom" for that specific sink model.
Genie generates a quick interactive "visual check" for the tenant: "Can you confirm the leak is coming from the point shown in this 3D model?"

This reduces "false alarm" maintenance calls by up to 40%.

Would you like me to create a "Property Policy" prompt that helps Ultra turn an apartment's House Rules into a RAG-ready Knowledge Atom?

From <https://gemini.google.com/app/> Google Gemini 3.0 Pro (Thinking)