Google DeepMind on May 19, 2026 introduced Gemini 3.5, a new model family positioned around agentic execution rather than purely conversational capability. The first release in the series, Gemini 3.5 Flash, is now broadly available across consumer, developer, and enterprise surfaces.

Performance Claims

According to DeepMind, 3.5 Flash outperforms Gemini 3.1 Pro on several agentic and coding benchmarks, including Terminal-Bench 2.1 (76.2%), GDPval-AA (1656 Elo), and MCP Atlas (83.6%). The model also posts an 84.2% score on CharXiv Reasoning, a multimodal understanding benchmark. DeepMind states the model runs at roughly four times the output-token throughput of other current frontier models, and places it in the top tier of the Artificial Analysis intelligence-versus-speed index.

Agentic Architecture and the Antigravity Harness

A notable architectural element is the pairing of 3.5 Flash with Google Antigravity, described as an agent-first development platform. This harness enables the model to coordinate collaborative subagents across multi-step workflows. Cited examples include autonomously migrating a legacy codebase to Next.js, synthesizing a research paper and producing a playable game implementation in approximately six hours using two cooperative agents, and dynamically renaming and categorizing unstructured assets.

DeepMind frames the speed and cost profile, claimed at less than half the cost of competing frontier models for comparable tasks, as enabling agentic use cases that were previously impractical at scale.

Enterprise Deployments

Several organizations are already running the model in production or pilot contexts:

  • Macquarie Bank is piloting the model for customer onboarding, using it to reason over documents exceeding 100 pages and surface relevant recommendations with low latency.
  • Salesforce is integrating 3.5 Flash into its Agentforce product to automate enterprise tasks via multi-subagent deployments with persistent context and multi-turn tool calling.
  • Xero is using agents to autonomously manage multi-week workflows such as supplier identification and 1099 tax form preparation.
  • Databricks is applying agentic workflows to real-time monitoring, dataset reasoning, and automated issue diagnosis for data science teams.
  • Ramp is leveraging the model’s multimodal capabilities for improved invoice OCR combined with historical-pattern reasoning.
  • Shopify is running parallel subagents to analyze complex data for merchant growth forecasting at global scale.

Consumer Rollout and Gemini Spark

3.5 Flash is now the default model powering the Gemini app and Google Search’s AI Mode globally. DeepMind also announced Gemini Spark, a personal AI agent built on 3.5 Flash that is intended to run continuously, taking action on behalf of users under their direction. A trusted-tester rollout began on the day of announcement.

What Comes Next

Gemini 3.5 Pro is described as already in internal use at Google, with a public rollout planned for the following month. DeepMind offered no additional benchmark or capability details for Pro at this stage.

For security and infrastructure teams, the primary relevance of this release is the expanded capacity of AI agents to autonomously execute multi-step workflows, access external tools, and operate across enterprise data environments, areas that raise ongoing questions around access control, audit trails, and prompt-injection risk in agentic pipelines.