Google splits its new Gemini lineup into a creation model and an agent model

Two models with deliberately different jobs

Google announced two distinct things and was careful not to blur them. Gemini Omni is framed as a creation model — it takes images, audio, video, and text as input and produces video, with conversational editing across turns. Gemini 3.5 is framed as an execution family, opening with 3.5 Flash, aimed at agents and coding on long-horizon tasks.

With Gemini Omni, Gemini's ability to reason meets the ability to create, while Gemini 3.5 is built to help you execute complex, agentic workflows.Montana Labs

The Omni demos all revolve around one property: continuity. The source describes edits where "characters stay consistent, the physics hold up and the scene remembers what came before." The violinist carousel — playing a song, then transported to a new environment, then with the violin made invisible, then re-angled over the shoulder — is a demonstration of state persistence across instructions, not one-shot generation.

Antigravity is the piece appearing under everything

The most repeated name in the 3.5 Flash section isn't the model — it's Antigravity, described as a harness. It shows up in the asset-renaming demo, in the collaborative subagents pitch for "the most demanding use cases," inside Search's ability to build custom experiences, and as the runtime for Gemini Spark.

That repetition matters. Google is positioning 3.5 Flash and Antigravity as a paired stack: the model supplies frontier-level reasoning at Flash speeds, and the harness supplies the orchestration for multi-step work under supervision. The source is explicit that reliable execution happens "under supervision" — a hedge worth noting for anyone evaluating these for production.

Agentic features are being pushed directly into Search

3.5 Flash is now the default model for both the Gemini app and AI Mode in Search globally. On top of that, Google is embedding several agentic behaviors into Search itself: information agents that run 24/7 in the background and send comprehensive updates with links; generative UI that builds a response "in the right format for your question — completely on the fly"; and custom experiences like dashboards, trackers, or mini apps that persist across sessions.

The examples are concrete — an information agent tracking sneaker collabs from a user's favorite athletes, Search building an interactive Gyroid-pattern visual, Search assembling a custom fitness tracker. This is Google using its highest-traffic surface to demonstrate that generation and agents now happen inside the results, not in a separate product.

The rollout map tells you who gets what, and when

The distribution details are staggered by tier and geography. Generative UI in Search is promised for everyone, free, this summer. Information agents launch first for AI Pro and Ultra subscribers this summer. Building your own custom experiences with Antigravity in Search comes in the following months, starting with AI Pro and Ultra in the U.S. Gemini Spark — the personal agent integrated with Gmail, Docs, and Slides — is available now, but only to Google AI Ultra subscribers in the U.S.

Omni Flash is rolling out to AI Plus, Pro, and Ultra subscribers globally through the Gemini app and Google Flow, free on YouTube Shorts and the YouTube Create App, with developer and enterprise API access coming in the weeks after. 3.5 Flash is generally available now across Antigravity, the Gemini API in AI Studio and Android Studio, the Gemini Enterprise Agent Platform, and Gemini Enterprise.

What the split means for teams building on Gemini

For applied teams, the practical takeaway is that Google has separated the buying decisions. If you need video creation, that's Omni, and it arrives via consumer apps first with APIs trailing by weeks. If you need agentic execution and coding, that's 3.5 Flash, and it's already generally available through developer and enterprise channels alongside the Antigravity harness.

The most consequential detail is that the flagship agent experiences — Spark, information agents, in-Search custom apps — are gated behind Ultra and Pro tiers and, in several cases, the U.S. only at launch. The free surfaces (default 3.5 Flash in Search and the Gemini app, generative UI in Search this summer) get the model but not the full agent stack. Anyone planning to depend on Antigravity-driven autonomy should also note Google's own framing that this runs "under your direction" and "under supervision" — the automation is deliberately kept on a leash.