At its developer event Google I/O, Google announced Gemini 3.5, its new family of AI models[1]. As the first release in the series, the company has begun rolling out "3.5 Flash," which combines high speed with strong processing power, to users around the world[1]. Available to anyone through the Gemini app and AI Mode in Search, it is a notable update that the company says significantly raises the model's performance as an agent that carries out tasks on its own.

3.5 Flash, the First Model in the Lineup

Gemini 3.5 is Google's latest family of models, which the company describes as combining "frontier intelligence with action"[1]. The first release in the series is 3.5 Flash, part of the lightweight, high-speed "Flash" line[1]. According to Google, 3.5 Flash is already available to billions of people worldwide through the Gemini app and AI Mode in Google Search. For developers it is offered through the agent-first development platform Google Antigravity and the Gemini API in Google AI Studio and Android Studio, and for enterprises through the Gemini Enterprise Agent Platform and Gemini Enterprise[1].

The higher-end "3.5 Pro" is also in development; it is already in use internally, and the company says it plans to begin rolling it out the following month[1].

Benchmarks Point to "Speed and Intelligence Together"

Google says 3.5 Flash delivers intelligence rivaling that of large flagship models, at the speed expected of the Flash series[1]. By the company's own figures, it outperforms the higher-tier Gemini 3.1 Pro on coding and agentic benchmarks, scoring 76.2 percent on Terminal-Bench 2.1, 83.6 percent on MCP Atlas, and 84.2 percent on CharXiv Reasoning, which measures multimodal understanding[1]. On GDPval-AA, which evaluates agentic ability, it reached 1656 Elo[1].

The company also states that in output tokens per second it achieves speeds four times those of other frontier models, emphasizing that users no longer have to give up either quality or response speed[1]. It is worth keeping in mind that all of these figures are published by Google itself.

Agentic Capabilities That Handle Long-Horizon Tasks

3.5 Flash is said to be especially strong at "long-horizon tasks" that involve many steps and can take days or weeks[1]. Google cites examples of work that used to take developers days or auditors weeks now being completed in a fraction of the time, often at less than half the cost of other frontier models[1].

Combined with the Antigravity development platform, it can link multiple subagents to tackle large-scale problems, the company says[1]. In published demos, two agents read and interpreted the AlphaZero paper to build a fully playable game in six hours, and another example transformed a messy legacy codebase into Next.js[1].

Enterprise Adoption and the Personal Agent "Gemini Spark"

Real-world adoption is already underway. According to Google, Shopify runs multiple subagents in parallel to produce more accurate growth forecasts for merchants at a global scale[1]. Macquarie Bank is piloting how the model can read documents of more than 100 pages to speed up customer onboarding, and Salesforce has built 3.5 Flash into Agentforce to automate complex enterprise tasks[1]. Ramp, Xero, and Databricks were also cited as beginning to use agents for work such as reading invoices, preparing tax documents, and analyzing data[1].

For consumers, the new personal AI agent "Gemini Spark" uses 3.5 Flash[1]. It runs around the clock to support users' digital lives, carrying out tasks on their behalf under their direction[1]. Google has begun offering it to trusted testers and says it plans to bring a beta to Google AI Ultra subscribers in the US the following week[1]. 3.5 Flash has also become the default model for the Gemini app and AI Mode in Google Search worldwide[1].

Built with Frontier Safeguards

Gemini 3.5 was developed in line with Google's Frontier Safety Framework[1]. The company says it has strengthened safeguards related to cyber attacks and CBRN (chemical, biological, radiological, and nuclear) risks, making the model less likely to generate harmful content while also reducing excessive refusals of safe queries[1]. It says it has incorporated interpretability tools that help check and understand the AI's internal reasoning before it returns an answer[1].

Summary

Gemini 3.5 is a generation that puts an agent's practical usefulness front and center, under the banner of combining speed and intelligence. The first release, 3.5 Flash, can be used right away through the Gemini app and Search, and has also been broadly rolled out for enterprises and developers. Many of the performance figures cited are Google's own, but with 3.5 Pro due the following month, its real-world capabilities will be worth watching.

出典:https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5/