Google Gemini 3.5 Flash & Omni: Agentic AI Models at I/O 2026

TL;DR: At I/O 2026 Google launched Gemini 3.5 Flash—a fast, low-cost frontier model—and Gemini Omni, a fully multimodal creator, alongside Antigravity 2.0, a developer-focused agent platform. The tools target long-horizon tasks, enterprise productivity, and scientific research, signaling a move from assistance to autonomous AI agents.

Key facts

• Gemini 3.5 Flash, the first model that blends frontier-level intelligence with actionable speed, is generally available through Antigravity, the Gemini API, Google AI Studio and Android Studio [3][8].
• On benchmarks Gemini 3.5 Flash outperforms Gemini 3.1 Pro: Terminal-Bench 2.1 (76.2%), GDPval-AA (1,656 Elo) and MCP Atlas (83.6%) [3].
• The model runs at less than half the cost of comparable frontier models, making it suitable for long-horizon agentic tasks such as rapid code development and complex UI generation [3].
• Gemini Omni expands multimodal capability, allowing creation from any input—starting with video—and is billed as a leap forward in world understanding, editing and generative flexibility [2][3][7].
• Antigravity 2.0 adds a CLI, SDK and managed-agents API, letting developers build, sandbox and host custom agents on their own infrastructure [2][8].
• Gemini Spark provides a 24/7 personal AI agent for Google Workspace that can autonomously monitor inboxes and draft replies [6][7].
• Gemini for Science bundles Empirical Research Assistance (ERA) and Co-Scientist, both published in Nature, and open-sourced on GitHub, accelerating projects from neuroscience to cosmology [1].

Google’s Agentic Turn at I/O 2026

At the May 2026 developer conference, Google framed its latest research releases as the start of a new “agentic era.” The headline announcements were two next-generation Gemini models—Gemini 3.5 Flash and Gemini Omni—and a major upgrade to the Antigravity developer platform, now called Antigravity 2.0. Together they move Google’s AI strategy from assistant-style tools toward autonomous agents that can carry out complex, long-running workflows.

Gemini 3.5 Flash: Speed Meets Frontier Intelligence

Gemini 3.5 Flash is billed as the first model that mixes frontier-level reasoning with actionable speed. It is now generally available through four channels: the Antigravity platform, the Gemini API, Google AI Studio, and Android Studio [3][8].

In head-to-head benchmark tests the model outperformed its predecessor, Gemini 3.1 Pro, on three well-known metrics:

• Terminal-Bench 2.1 – 76.2% accuracy
• GDPval-AA – 1,656 Elo rating
• MCP Atlas – 83.6% score

These numbers place Gemini 3.5 Flash in the top-right quadrant of the Artificial Analysis index, indicating both high quality and low latency [3].

Cost is another differentiator. Google says the model runs at less than half the price of comparable frontier models, a claim that directly addresses the expense barrier that has limited large-scale agentic deployments [3]. The company envisions the model handling “long-horizon” tasks such as rapid code generation, financial document preparation, and complex UI creation.

Gemini Omni: Full-Spectrum Multimodality

While Flash focuses on speed and cost, Gemini Omni expands the creative canvas. The model can accept any input type—including raw video—and generate any output, from text to images to code. Google describes this as a “leap forward in world understanding, multimodality and editing” [2][3][7].

The multimodal capability is already being woven into consumer products. The Gemini app now offers Gemini Spark (a persistent personal assistant) and Daily Brief (a concise news digest). Other integrations include Google Pics, intelligent eyewear, Ask YouTube, and the new Universal Cart shopping experience, where Omni powers real-time product recommendations based on visual cues.

Antigravity 2.0: An Agent-First Development Platform

The hardware and software that let developers build these agents received a substantial upgrade. Antigravity 2.0 adds a command-line interface (CLI), a software development kit (SDK), and a managed-agents API. These tools let developers spin up specialized sub-agents, run them safely in sandboxes, and even host custom agents on their own infrastructure [2][8].

The platform’s flexibility is already evident in several Google products. Information agents built on Antigravity now appear in Search, the Gemini app (as Spark and Daily Brief), and the Universal Cart intelligent shopping cart [2][8]. For enterprise customers, Google Cloud introduced the Gemini Enterprise Agent Platform, delivering both Gemini 3.5 Flash and Gemini Omni directly to business users, alongside Gemini Spark, a 24/7 personal AI agent for Google Workspace that can autonomously monitor inboxes and draft replies [6][7].

Gemini for Science: Accelerating Research

Google’s research arm highlighted a separate line of announcements aimed at scientific discovery. Gemini for Science bundles two tools—Empirical Research Assistance (ERA) and Co-Scientist—both of which were recently described in Nature [1].

ERA helps scientists write expert-level empirical code, while Co-Scientist assists with data analysis and hypothesis generation. Early adopters report accelerated progress in fields ranging from neuroscience to cosmology. Notable case studies include:

• Predicting hospital admissions for respiratory illnesses
• Forecasting seasonal runoff in California river basins

All code is open-sourced on GitHub, allowing the global research community to extend and adapt the tools [1].

Scale of Adoption: Token Processing Numbers

During the keynote, Sundar Pichai highlighted the sheer scale of AI usage across Google’s ecosystem. Token processing—essentially the amount of text the models handle—has risen from 9.7 trillion per month two years ago to over 3.2 quadrillion per month, a seven-fold increase since the previous I/O event [4]. This growth underscores why faster, cheaper models like Gemini 3.5 Flash are critical for keeping costs manageable while supporting ever-more ambitious agentic applications.

Implications for Southeast Asian Developers

While the announcements are global, they have particular relevance for developers in Thailand and the broader Southeast Asian region. The open-source nature of Gemini for Science and the availability of Antigravity 2.0 through Google Cloud lower the barrier for local startups and university labs to experiment with autonomous agents. Industry observers note that the agentic paradigm could reshape how Thai companies automate customer support, build intelligent IoT solutions for smart homes, and accelerate scientific collaborations across the region. However, concrete adoption metrics will emerge only as the tools mature and local ecosystems build the necessary expertise.

Looking Ahead

Google’s I/O 2026 reveals a clear strategic shift: from AI that merely assists to AI that can act independently. By pairing powerful, cost-effective models with a robust agent-first platform, Google is positioning itself to drive the next wave of productivity gains in both consumer and enterprise contexts, while also empowering scientific discovery. The success of this vision will depend on how quickly developers—inside and outside Google—can harness these tools to create trustworthy, reliable agents that truly augment human ingenuity.

Sources Used: [1][2][3][4][5][6][7][8]

Sources

3. 100 things we announced at I/O 2026 (blog.google) — 2026-05-20
4. All the news from the Google I/O 2026 Developer keynote (developers.googleblog.com) — 2026-05-19
5. I/O 2026 (blog.google) — 2026-05-19
6. Google I/O 2026: What’s New and How You Can Actually Use It (www.linkedin.com) — 2026-05-22
7. Innovations from Google I/O 26 on Google Cloud | Google Cloud Blog (cloud.google.com) — 2026-05-20
8. A New Era of Discovery: Google Research at I/O 2026 (research.google) — 2026-05-28
9. I/O 2026: Welcome to the agentic Gemini era (blog.google) — 2026-05-19
10. Google I/O 2026 | Google Home Developers (developers.home.google.com) — 2026-01-01