Gemini 3.5 family

What it is

The Gemini 3.5 family is the generation of large language models Google DeepMind unveiled for its I/O developer conference on 19 May 2026. The generation opened with Gemini 3.5 Flash, a generally available model built for agentic and coding work, and Google moved quickly to put it at the centre of its products, making it the default model in the Gemini app and in AI Mode in Search worldwide. On 21 July 2026 the workhorse seat changed hands again when Google launched Gemini 3.6 Flash; a Gemini 3.5 Pro remains in partner testing rather than generally available, so the heavier tiers have yet to move to the new generation.

Flash as the default

Making Flash the default is a significant decision. AI Mode in Search reaches an enormous audience, so the model chosen to answer those queries is used far more often than any specialist tier. By putting Gemini 3.5 Flash in that seat, Google has signalled that its newest, agent-focused model is fast and cheap enough to serve at the scale of Search, rather than being held back for premium users. “Flash” in Google’s naming denotes the quick, cost-efficient tier, which makes its promotion to the default all the more telling: the everyday model most people meet is now a current-generation one.

The wider lineup

Around the new Flash sits a wider lineup that shows how Google now segments the range. Gemini 3.1 Pro, released on 19 February 2026, holds the Pro tier for heavier general work. Gemini 3 Deep Think, upgraded on 12 February 2026, is aimed at frontier reasoning, the hardest multi-step problems where extra deliberation pays off. Gemini 3.5 Flash-Lite, generally available since 21 July 2026, sits at the bottom for the most cost-sensitive tasks. The result is a ladder running from cheap and fast to slow and powerful, with the Gemini 3.5 line the newest rung.

The version numbers repay a moment’s attention. The 3.5 label denotes the newest wave, while Gemini 3.1 Pro, which still holds the Pro tier, sits a half-step behind, and Deep Think stands apart as the reasoning specialist. Google is iterating on parts of the range at different speeds, refreshing the high-volume Flash tier first and letting the heavier tiers follow. For most users the practical question is simply which model answers their query, and increasingly the answer is a 3.5-generation one.

Tools and pricing

Google has paired the models with tooling aimed at developers building agents. Google Antigravity is an agentic development environment, a workspace for building and running software agents. Managed Agents in the Gemini API let developers hand the plumbing of an agent to Google’s infrastructure rather than wiring it up themselves. The emphasis on agentic and coding work runs through both the model design and the tools, which fits the direction the whole field has taken this year. Readers following that shift can track it on our large language models hub.

On the commercial side, Google AI Ultra, the top consumer subscription, is now available from $100 per month. That anchors the premium end of the range for individuals, while the Gemini API and the default-in-Search placement cover developers and the mass market respectively. Taken together it is a familiar three-way split: a flagship subscription, a developer API, and a free tier reaching a vast audience through Search. Charging a three-figure monthly sum for the top tier only works if the flagship reasoning and agentic features clearly justify it.

What to watch

Where the Gemini 3.5 family sits in the wider field is as Google’s answer to a summer of frontier launches from rival labs. Its distinguishing move is distribution: by making a current-generation model the default in Search and the Gemini app, Google reaches a scale no competitor can match, whatever the benchmark tables say. The Flash seat has since passed to Gemini 3.6 Flash (21 July 2026), while Google has said a broader 3.5 Pro rollout is still in testing with partners. What to watch next is when the Pro tier catches up, and whether the agentic tooling becomes a developer default. More context sits on our main AI hub.