Google: Gemini 3.5 Flash and Pro — the fastest frontier models yet

Google I/O 2026: Gemini 3.5 Flash available now, Pro coming soon

At the Google I/O 2026 conference, Google unveiled Gemini 3.5 Flash, a new frontier model available today in the Gemini app, Google Search, and the Gemini API. Its sibling, Gemini 3.5 Pro, is currently in internal testing and entering broader rollout next month.

The key claim: Gemini 3.5 Flash generates output tokens four times faster than other frontier models, with inference costs that Google says are less than half those of competing solutions. On the Terminal-Bench 2.1 benchmark the model achieves 76.2%, surpassing even Gemini 3.1 Pro, while on the GDPval-AA agentic evaluation it scores 1,656 Elo points.

What is agentic AI and why is it now a priority?

Agentic AI refers to systems that autonomously plan and execute a sequence of actions without constant user supervision — searching the web, writing code, calling external services, and making decisions within a single task. Unlike classical chatbots that answer individual questions, an agentic model can manage a project through dozens of steps.

Gemini 3.5 Flash is designed exactly for such scenarios. High execution speed is critical when a model must make hundreds of small calls in a short time, and an error in one step can be quickly compensated by a retry without a long wait.

Gemini Spark: a personal AI agent running 24 hours a day

Gemini Spark is a new type of product — a persistent personal AI agent based on Gemini 3.5 Flash that continuously monitors and organizes users’ digital lives. The agent can schedule meetings, summarize emails, track deadlines, and proactively alert users to changes without any manual prompting.

Spark is currently being tested with selected users, and the beta arrives next week for Google AI Ultra plan subscribers in the US. Broader global availability is expected by late summer 2026.

Antigravity 2.0 and new TPU chips

Antigravity 2.0 is Google’s developer platform for building agentic applications. The upgrade brings parallel sub-agent execution, meaning a single user request can simultaneously launch multiple specialized agents working in coordination. The platform is available through the Gemini API, AI Studio, and Android Studio.

Underpinning all of this are the new eighth-generation TPU v8t and v8i chips — TPU v8t is optimized for training large models, and TPU v8i for fast inference. The combination of these chips explains why Gemini 3.5 Flash can deliver 4× speed without compromising accuracy.

Partner companies already testing Gemini 3.5 Flash in production include Shopify, Salesforce, Macquarie Bank, and Databricks — suggesting the model is ready for enterprise use beyond lab conditions.