Skip to content

Gemini 3.5: Frontier Intelligence with Action

Overview

Gemini 3.5 is Google DeepMind's third-generation model series, announced at Google I/O on May 19, 2026. The series is explicitly architected for agentic workflows — multi-step, long-horizon tasks that require the model to plan, act, and iterate across many tool calls, often in parallel. The headline claim is that the Flash variant in this generation matches or exceeds large flagship models on several dimensions while running at significantly higher speed, positioning the 3.5 series as a step-change rather than an incremental improvement.

The series launches with 3.5 Flash immediately generally available. 3.5 Pro is in active internal use at Google and is planned for public rollout in June 2026.

3.5 Flash: Capabilities and Benchmarks

At launch, 3.5 Flash surpasses Gemini 3.1 Pro on coding and agentic benchmarks — meaning the Flash variant in the 3.5 generation exceeds the previous generation's Pro variant. Benchmark results as of May 19, 2026:

  • Terminal-Bench 2.1: 76.2% — command-line task completion in agentic settings
  • GDPval-AA: 1,656 Elo — agentic task evaluation rating
  • MCP Atlas: 83.6% — tool-calling capability across Model Context Protocol scenarios
  • CharXiv Reasoning: 84.2% — multimodal reasoning over charts and figures

On the Artificial Analysis intelligence-speed index, 3.5 Flash delivers 4× higher output tokens per second compared to other frontier models, placing it in what Google describes as the top-right quadrant of the intelligence-speed tradeoff. This challenges the conventional assumption that frontier-level capability requires frontier-level latency.

Agentic Architecture: The Antigravity Integration

3.5 Flash is designed to run natively inside the Antigravity agent harness, Google's execution framework for multi-agent systems. Under Antigravity, the model:

  • Deploys and coordinates collaborative subagents in parallel, each working on independent subtasks
  • Sustains frontier performance across long-horizon multi-step workflows without degradation
  • Executes complex multi-turn tool calling with reliable context retention across turns
  • Iterates on code, assets, and documents in a self-improvement loop

Demonstrations at I/O 2026 included: synthesizing an academic paper and coding a fully playable game in six hours (two-agent builder-player loop); transforming a legacy codebase to Next.js; generating city landscapes in Antigravity; developing a game through rapid self-improvement. These examples represent classes of work — sustained multi-step creative and engineering tasks — that earlier agentic systems completed unreliably or not at all in a single session.

UI and Multimodal Generation

Building on the multimodal foundation of Gemini 3, 3.5 Flash generates richer and more interactive web UIs and graphics. Use cases demonstrated at I/O 2026:

  • Creating interactive animations directly from research papers
  • Converting plain text descriptions into interactive hardware prototypes
  • Executing multiple branding concepts in parallel from a single prompt
  • Generating different UX approaches for a checkout flow in under 60 seconds

The underlying capability is the model's ability to interpret visual specifications, reason about interactive behaviour, and output implementation-ready code in a single pass — reducing the round-trips between specification, design, and engineering.

3.5 Pro

3.5 Pro is positioned as the high-capability variant for maximum-complexity tasks. At the Google I/O announcement, it was in active internal use at Google DeepMind. Public availability is planned for June 2026. Google has not released 3.5 Pro benchmarks alongside the Flash launch.

Gemini Spark: Personal AI Agent

Gemini Spark is a new personal AI agent product powered by 3.5 Flash. It runs continuously, takes action on the user's behalf, and is designed to navigate a user's digital life under their supervision — managing tasks, communications, and workflows autonomously. Gemini Spark differs from the Gemini app in that it operates proactively, initiating actions rather than waiting for queries.

At I/O 2026, Google began rolling out Spark to trusted testers immediately, with a Beta planned for Google AI Ultra subscribers in the US the following week. 3.5 Flash also became the default model for the Gemini app and AI Mode in Google Search globally.

Enterprise Partner Deployments

Multiple enterprise partners demonstrated production use of 3.5 Flash at I/O 2026, providing early evidence of the model's production agentic characteristics:

  • Shopify: Parallel subagents for long-horizon merchant growth forecast analysis at global scale.
  • Macquarie Bank: Customer onboarding acceleration — the model reasons over complex 100+ page documents, retrieves relevant information, and makes low-latency recommendations.
  • Salesforce: Integration into Agentforce for complicated enterprise task automation via multi-subagent deployment with context retention and complex multi-turn tool calling.
  • Ramp: Smarter OCR through multimodal understanding of complex invoices combined with reasoning over historical patterns.
  • Xero: Autonomous management of multi-week workflows such as identifying suppliers and gathering 1099 tax form information for small businesses.
  • Databricks: Agentic workflows for real-time monitoring and reasoning across massive datasets to diagnose issues, identify fixes, and propose solutions.

This breadth of deployment contexts — financial services, accounting, e-commerce, SaaS — suggests the model generalises across industry-specific document and workflow structures rather than excelling in only narrow domains.

Safety and Governance

3.5 Flash was developed under Google's Frontier Safety Framework. Specific improvements in this generation include stronger cyber and CBRN (chemical, biological, radiological, nuclear) safeguards that reduce the likelihood of harmful content generation, along with reduced false refusals on safe queries. A notable addition is new interpretability tooling that inspects and understands the model's inner reasoning before it produces a response — used during safety training and as a mitigation layer in production.

Availability

3.5 Flash is immediately available via: - Google Antigravity — the agent development platform - Gemini API in Google AI Studio and Android Studio - Gemini Enterprise Agent Platform and Gemini Enterprise - Gemini app (global default model) - AI Mode in Google Search (globally)

Competitive Notes

The 3.5 Flash vs Pro naming convention breaks from the prior generation's convention: in the 3.5 generation, Flash matches or exceeds 3.1 Pro on key benchmarks, which represents a generational uplift rather than a tier compromise. Direct comparisons with GPT-5.5 and Claude Opus 4.7 on the same benchmarks were not published at launch. GitHub Copilot's integration timeline for Gemini 3.5 models has not been announced as of the I/O date.

Changelog

2026-05-20 — Page created from blog.google I/O 2026 announcement, May 19, 2026 (Type A, confidence 78 — single official primary source)