Google introduced Gemini 3.5, a new model family that combines advanced intelligence with the ability to take action. The company began the rollout with Gemini 3.5 Flash, a model designed for AI agents, programming, and complex multistep tasks.
The promise is straightforward: deliver performance comparable to that of large models, but with the speed characteristic of the Flash series. What does that mean in practice? An agent can plan, use tools, write code, and correct its own results without making you wait as long between one instruction and the next.
Gemini 3.5 Flash bets on speed and action
According to Google, Gemini 3.5 Flash outperforms Gemini 3.1 Pro on several benchmarks related to programming and agents. The announced results include:
- 76.2% on Terminal-Bench 2.1, a test focused on tasks performed from the terminal.
- 1,656 Elo points on GDPval-AA, a performance evaluation for professional work.
- 83.6% on MCP Atlas, focused on tool use and agentic capabilities.
- 84.2% on CharXiv Reasoning, which measures multimodal understanding and reasoning about charts.
Google also says the model generates tokens up to four times faster than other frontier models. Token generation speed matters because it determines how long the system takes to produce a response or complete a sequence of actions.
The central idea behind Gemini 3.5 Flash is to reduce the trade-off between quality and latency: respond with advanced intelligence without forcing users to wait too long.
These figures come from evaluations reported by Google. As with any benchmark, it is worth interpreting them alongside independent testing and results from real-world scenarios.
Agents capable of completing long tasks
Gemini 3.5 Flash is designed for long-horizon tasks—that is, processes that require planning, execution, review, and several consecutive steps. A developer could use it to create an application, modify a codebase, and verify that the changes work.
In a financial company, the model could help prepare documents or automate parts of an audit workflow. In a data science team, it could explore large volumes of information and find patterns that professionals would then need to review.
Google says these tasks can be completed in a fraction of the time and, in some cases, for less than half the cost of other frontier models. However, the actual savings will depend on usage volume, the connected tools, and the level of human supervision.
Subagents and coordinated work
When combined with the updated version of Antigravity, Gemini 3.5 Flash can coordinate collaborative subagents. Each subagent can handle one part of the problem while the main system organizes the process.
For example, one agent could analyze an application's requirements, another could write the code, a third could run tests, and a fourth could check for potential errors. Supervision is still necessary, especially when actions affect data, money, or production systems.
The model also takes advantage of Gemini 3's multimodal foundation to create more interactive web interfaces and graphics. This expands its usefulness beyond text and code: it can interpret visual information and help build complete digital experiences.
Availability for users, developers, and businesses
Gemini 3.5 Flash is available globally to billions of people through:
- The Gemini app.
- AI Mode in Google Search.
- Google Antigravity.
- The Gemini API in Google AI Studio and Android Studio.
- Gemini Enterprise Agent Platform.
- Gemini Enterprise.
Google also made Gemini 3.5 Flash the default model for the Gemini app and AI Mode in Search worldwide.
For developers, this means they can test its capabilities from environments for building agents and applications. For businesses, the appeal lies in automating internal processes that previously required weeks of manual work, although integration should be accompanied by access controls, activity logs, and result reviews.
Gemini Spark brings agents into personal use
One of the new features presented is Gemini Spark, a personal agent based on Gemini 3.5 Flash. Google describes it as a system capable of running continuously and helping manage a user's digital life under their direction.
The company has begun testing it with trusted users and plans to bring a beta version to Google AI Ultra subscribers in the United States. The concept is reminiscent of an assistant that not only answers questions but can also take authorized actions on a person's behalf.
That shift matters. A chatbot waits for an instruction and provides a response. An agent tries to complete a goal by using tools and making intermediate decisions. That is why authorization, transparency, and the ability to stop a task will be just as important as the model's capabilities.
Frontier safety and safeguards
Google says Gemini 3.5 was developed under its Frontier Safety Framework, with improved protections against harmful uses related to cybersecurity and chemical, biological, radiological, or nuclear materials.
The company also says the model seeks to reduce two opposing problems: generating dangerous content and incorrectly rejecting legitimate requests. To do this, it incorporated more advanced safety training, new mitigations, and interpretability tools that help review its internal reasoning before delivering a response.
This does not eliminate every risk. An agent with access to tools can have a greater impact than a model that only generates text, especially if it has permission to modify files, send messages, or interact with external services. Human supervision and authorization limits remain essential.
A step toward more useful agents
Gemini 3.5 Flash shows where the AI competition is heading: it is no longer only about answering better, but also about planning and acting within real workflows. Speed, cost, and connections to tools are now just as important as conversation quality.
Google also confirmed that it is working on Gemini 3.5 Pro, which is already being used internally and is expected to launch next month. In the meantime, Flash aims to become the practical option for those who need fast agents, assisted programming, and automation at scale.
The question is no longer whether AI can perform an isolated task. The question is how much control we want to give it when it has to complete the entire process for us.
Original source
https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-5