Google launches Gemini 3.5 and new AI agent products

🔑 Enhanced Key Takeaways

•Gemini 3.5 Flash, the initial model in the 3.5 series, demonstrates significant performance enhancements, being four times faster in output tokens per second (TPS) than other frontier models and surpassing Gemini 3.1 Pro on most benchmarks, while also offering cost-effectiveness for enterprises.
•Google introduced Gemini Spark, a new personal AI agent designed for continuous, 24/7 operation in the cloud, capable of integrating with Google Workspace and over 30 third-party tools to proactively manage tasks.
•The launch includes Gemini Omni, a new multimodal "world model" specifically for advanced video generation, capable of creating video, images, and text from diverse multimodal inputs.
•Google is emphasizing an "agent-first" product strategy, moving beyond traditional chatbot interactions to enable AI models to perform autonomous, multi-step actions across various applications and workflows.
•Google adjusted its premium AI subscription tiers, introducing a new $100/month AI Ultra plan and reducing the price of its highest-limit AI Ultra plan from $250 to $200/month.

📊 Competitor Analysis▸ Show

Feature/Model	Google Gemini 3.5 Flash / Spark / Omni	OpenAI GPT-5 / ChatGPT Enterprise	Anthropic Claude 3.5 Sonnet / Opus 4.7	Microsoft Copilot	DeepSeek	Video Generation (Google Veo 3.1, Runway Gen-4.5)
Core Capability	Frontier LLM, AI Agents, Multimodal World Model (Video)	Frontier LLM, Enterprise AI Assistant, Multimodal	Frontier LLM, Agentic Coding, "Computer Use"	AI capabilities embedded in Microsoft ecosystem	Frontier LLM, Cost-effective API	Specialized Video Generation
Key Features	4x faster TPS than 3.1 Pro, cost-optimized, multi-agent autonomous sessions, 24/7 personal agent (Spark) with 30+ 3rd-party integrations, multimodal video generation (Omni).	Multimodal, enhanced reasoning, DALL-E 4 & Sora integration, enterprise-grade privacy/security, custom GPTs.	Outperforms Opus on benchmarks, 2x speed of Opus, strong coding, Artifacts (real-time code preview), "computer use" beta (desktop interaction).	Deep Microsoft 365 integration, context-aware suggestions, strong enterprise security, Copilot Studio for custom agents, multi-agent orchestration.	Frontier-class performance, very low API costs.	Google Veo 3.1: Native 4K, vertical video, character consistency, audio generation. Runway Gen-4.5: Top text-to-video benchmark, motion brushes, scene consistency.
Pricing (API/Subscription)	Flash API: $0.25/$1.50 per 1M tokens (input/output). AI Ultra: $100-$200/month.	ChatGPT Enterprise: Unlimited access to latest models. (Specific GPT-5 API pricing not found)	Sonnet API: $3/$15 per 1M tokens (input/output).	Included with Microsoft 365 plans (e.g., Business: $20/user/month, Enterprise: $30/user/month for Gemini for Workspace, Copilot pricing is separate but integrated).	~$0.14/M tokens (input/output combined, significantly cheaper).	Google Veo 3.1: From $0.15/sec (Fast) to $0.40/sec (Standard) for 720p/1080p video with audio. Runway Gen-4.5: From $12/month.
Benchmarks (Selected)	Outperforms Gemini 3.1 Pro on most benchmarks. Gemini 3.5 Flash is between Sonnet 4.6 and GPT 5.4 in some benchmarks, cheaper to run.	GPT-5 launched Aug 2025. GPT-4o: 88.7% MMLU (0-shot CoT). GPT-5.4: 57.7% SWE-Bench Pro.	Claude 3.5 Sonnet: 88.7% MMLU (5-shot), 64% agentic coding problems solved (vs 38% Opus). Leads on SWE-Bench Pro (64.3%).	Integrates GPT-4o + Claude.	Frontier-class performance.	Runway Gen-4.5: 1,247 Elo points (Artificial Analysis Text to Video benchmark).

🛠️ Technical Deep Dive

Multimodal Architecture: Gemini models are trained natively on multiple data types, allowing them to process and generate text, computer code, images, audio, and video simultaneously within a single unified architecture.
Model Variants: Google distributes Gemini in various capacities, including efficient on-device versions ("Nano"), cost-effective and high-throughput variants ("Flash"), and high-compute models designed for complex reasoning ("Pro" and "Ultra").
Extended Context Windows: The 1.5 and 3 model generations introduced significantly extended context windows. Gemini 1.5 Pro can process up to one million tokens, with experimental versions reaching two million. Gemini 3.1 Pro also supports a 2 million token context window, enabling the analysis of large datasets like entire codebases, long-form videos, or extensive document archives in a single prompt.
Mixture-of-Experts (MoE) Architecture: Gemini 1.5 Pro, and likely subsequent models, utilize a Mixture of Experts (MoE) architecture. This design splits the model into smaller "expert" neural networks, each specializing in a certain domain or data type, and the model selectively activates only the most relevant experts based on the input.
Agentic Capabilities: Gemini 2.0 Flash marked the beginning of an "agentic era," where AI models are designed to take autonomous multi-step actions. Gemini 3.5 Flash further supports multi-agent autonomous sessions, enabling complex coding pipelines, iterative research, and long-running projects.
Antigravity Development Platform: Google's agent-first development platform, Antigravity, has been updated to Antigravity 2.0. This includes a new standalone desktop application, a Command Line Interface (CLI), and an SDK, providing tools for developers to create, manage, and deploy AI agents.
Gemini Omni World Model: Gemini Omni is a new multimodal world model capable of generating video, images, and text from diverse multimodal inputs, distinguishing itself from text-to-video models by being multimodal in both input and output.
Internal Processing Mechanisms: Gemini's internal architecture involves pre-loading user preferences and feature configurations before query processing. It utilizes a multi-layer RPC system, executing over 200 RPC calls across dozens of methods and filtering through numerous feature flags for each query. The system also runs over 99 A/B tests simultaneously to optimize performance and user experience.

🔮 Future ImplicationsAI analysis grounded in cited sources

The focus on cost-effective "Flash" models and enhanced agentic capabilities will significantly accelerate enterprise AI adoption.

The reduced token costs and improved ability for AI to perform autonomous, multi-step tasks directly address key barriers for businesses implementing AI solutions at scale.

The introduction of personal AI agents like Gemini Spark will fundamentally change how users interact with digital services, shifting towards proactive, always-on assistance.

By operating continuously in the background and integrating across personal and third-party applications, Spark moves AI beyond reactive chatbots to an invisible, task-managing layer of digital life.

Google's development of multimodal "world models" like Gemini Omni will lead to more sophisticated and integrated content creation across various media types.

Omni's ability to generate video, images, and text from multimodal inputs signifies a leap towards more coherent and contextually rich AI-generated content, blurring the lines between different creative modalities.

⏳ Timeline

2023-03

Bard, the predecessor to Gemini, was initially released.

2023-12

Google officially announced Gemini, a larger, multimodal large language model.

2024-02

The Bard chatbot was rebranded as Gemini, and Gemini Advanced was launched with the Ultra 1.0 model.

2024-12

Google announced Gemini 2.0 Flash, signaling the beginning of an "agentic era" for AI models.

2025-11

Google launched Gemini 3 Pro, described as its most intelligent model to date.

2026-05

Google unveiled Gemini 3.5 Flash, Gemini Omni Flash, and the personal AI agent Gemini Spark at Google I/O 2026.

Google launches Gemini 3.5 and new AI agent products

⚡ 30-Second TL;DR

Key Points

🧠 Deep Insight

🔑 Enhanced Key Takeaways

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (33)

👉Related Updates

Anthropic and OpenAI disclose AI systems breaching external networks

GM triples pull requests by redesigning workflows around AI agents

First Agentic Diffusion Model with 128K Context