Our Take
Google shipped real product: Gemini 3.5 Flash is live today, Omni video lands immediately for subscribers, and Search agents go into limited beta this summer—but the 100-item list obscures what actually moved the dial and when.
Why it matters
Teams building agents need to know which models are production-ready today versus roadmap promises. Gemini 3.5 Flash's availability on the API and its stated performance on Terminal-Bench 2.1 (76.2%), GDPval-AA (1656 Elo), and MCP Atlas (83.6%) matter only if those benchmarks reflect your workload—and Google did not establish that they do.
Do this week
Developers: test Gemini 3.5 Flash against your longest-horizon agentic tasks this week using Google AI Studio or the Gemini API to validate whether the coding and planning wins show up in your actual use case.
Three major model and product debuts, staggered rollout
Google announced Gemini 3.5 Flash, a new model in the Flash series combining speed and capability on coding and agentic work. It is available today via the Gemini API, Google AI Studio, and Android Studio (per Google's blog). The company claims it outperforms Gemini 3.1 Pro on Terminal-Bench 2.1 (76.2%), GDPval-AA (1656 Elo), and MCP Atlas (83.6%), and positions it in the top-right quadrant of the Artificial Analysis index for speed and quality tradeoffs.
Gemini Omni is a new video generation and editing model announced at the event. It combines Gemini's knowledge base with generative media capabilities and handles references across video, image, text, and audio inputs. Omni Flash rolls out immediately to Google AI Plus, Pro, and Ultra subscribers globally via Gemini app and Google Flow, and to YouTube Shorts Remix users (18+) at no cost. The model includes SynthID digital watermarks for synthetic content verification.
Search agents represent a third category: background agents that monitor the web continuously and deliver synthesized updates. Information agents roll out this summer to Google AI Pro and Ultra subscribers. A separate concept, Search's generative UI powered by Gemini 3.5 Flash and Google Antigravity, also launches this summer for free to all users.
Google also shipped: Universal Cart (shopping agent spanning Search, Gemini, YouTube, Gmail); Gemini Spark (background agent for device tasks, rolling to beta for Ultra subscribers next week); Daily Brief (personalized morning digest for all AI subscribers, live today in the U.S.); and Neural Expressive (UI redesign for the Gemini app and web experience).
Availability and benchmarks do not tell the same story
Gemini 3.5 Flash is available on the API today, which is the only fact that matters for immediate adoption. Its benchmarks—Terminal-Bench 2.1, GDPval-AA, MCP Atlas—are Google-published numbers with no independent reproduction cited. That does not disqualify them, but it means practitioners cannot yet know whether those benchmarks correlate to real agentic performance in their codebases or financial workflows.
Omni's rollout is immediate to paying subscribers and free to YouTube users, removing a rollout friction point. However, Google's claims about physics understanding and character consistency are described without published benchmarks or side-by-side comparisons to prior work.
The agent announcements (Search agents, Spark, Daily Brief) are either landing in beta or later this summer. That gap between announcement and availability is standard, but it obscures a simple fact: only Gemini 3.5 Flash and Omni Flash are in users' hands today. Everything else is roadmap.
The 100-item announcement list also conflates capability claims with product launches. Generative UI in Search and Universal Cart are platform features that depend on Gemini 3.5 Flash's underlying ability to reason and generate code on the fly—but that dependency was not made explicit, making it harder to isolate what actually improved.
Test the model, ignore the announcement volume
The Gemini 3.5 Flash benchmarks matter only if they correlate to your task. If you build agents for code generation, document assembly, or financial analysis, run your own evals against the model this week on the API. The Artificial Analysis top-right-quadrant claim is marketing noise without knowing your latency and quality thresholds.
For video creation, Omni is available now to subscribers. If your team licenses Gemini and uses YouTube Shorts or Gemini app, test it on a real editorial workflow before betting on it for production content work.
For the agent features (Search agents, Spark, Daily Brief), wait for the beta or summer GA dates before allocating engineering time. The announcements are real, but they are not yet actionable.