Google introduces Gemini 3.5 Flash at I/O 2026

Background

Google I/O is Google's annual developer conference, and it is where the company presents new models and products to developers. On 19 May 2026 the conference was the setting for the announcement of the Gemini 3.5 family. The family did not appear all at once: it started with Gemini 3.5 Flash ^[1]. The naming is worth pausing on — the number marks a generation, and the first member of a generation to ship is usually the one most people end up using first.

What Google announced

The announcement was co-signed by Koray Kavukcuoglu, Jeff Dean, Oriol Vinyals and Noam Shazeer ^[1].

Gemini 3.5 Flash became the default model in the Gemini app and in AI Mode in Search ^[1]. Default is the operative word here. A default model is the one a product calls when the user makes no explicit choice, so whatever sits in that slot is the model most people meet without ever reading a model name. Putting a newly announced model into that slot in two products at once means the release reaches users through the app and through search, not only through a developer interface.

Alongside the model, Google introduced Gemini Spark, a personal agent that runs on 3.5 Flash ^[1]. An agent, in this sense, is a system that is given a goal and carries out steps toward it, rather than answering one question and stopping; a personal agent is one aimed at an individual user's own tasks. Google also introduced an updated Antigravity harness with collaborative subagents ^[1]. A harness is the software a model runs inside — the part that gives it tools, keeps track of where a task stands and decides what happens next — and subagents are smaller agents that a main agent hands parts of a job to, so that a task can be split up instead of handled in a single pass. "Collaborative" is the qualifier Google attached to them, which describes subagents working together on the same job.

The same day, Google released Antigravity 2.0 and the Antigravity CLI ^[3]. A CLI, or command-line interface, lets a developer drive a tool by typing commands in a terminal instead of clicking through a graphical window. Shipping a version number and a command-line tool on the same day suggests that the harness is being treated as a product with its own release schedule, not as an internal experiment.

The Gemini app itself gained two things: Gemini Omni video creation, and a new Google AI Ultra tier starting at $100 a month ^[2]. The first is a creation feature inside the consumer app; the second is a subscription tier, which sets a price for the highest level of access Google offers.

How to read the release

The items in the announcement fall into a few different layers. There is the model itself, Gemini 3.5 Flash, and the decision to make it the default in two large products ^[1]. There are tools for building with models: Gemini Spark as a packaged personal agent, and the Antigravity harness with collaborative subagents, released together with Antigravity 2.0 and the Antigravity CLI ^[1][3]. And there are features and pricing aimed at end users: Gemini Omni video creation and the Google AI Ultra tier starting at $100 a month ^[2].

That spread is what makes the date worth recording. A model release on its own is mostly a change for developers; a model release that simultaneously resets the default in the Gemini app and in AI Mode in Search, adds agent tooling, and adds a paid tier changes what ordinary users encounter and what they can pay for. The four co-signatories indicate that the research side of Google signed off on the release ^[1], while the same-day Antigravity 2.0 and CLI show the tooling side shipping on its own cadence ^[3].

One thing the announcement does not settle is how Gemini 3.5 Flash performs. The items here describe where the model is placed and what was released alongside it, not benchmark results or comparisons with other models. For a reader trying to follow the Gemini line over time, the useful anchors are the date, the name of the first model in the family, the products where it became the default, and the agent and tooling releases that came with it ^[1][2][3].

What it changed

Before this announcement, the Gemini 3.5 name did not exist in Google's lineup; after it, 3.5 Flash was the model behind the Gemini app and AI Mode in Search by default, whether or not a user asked for it ^[1]. The agent and harness releases — Gemini Spark, the updated Antigravity harness with collaborative subagents, Antigravity 2.0 and the Antigravity CLI — gave developers a set of tools built around the new model on the same day it was announced ^[1][3]. And the Gemini app's new video creation feature and the Google AI Ultra tier starting at $100 a month set out what the consumer side of the release looks like ^[2].