Google launches Gemini 3.5 Flash for agentic tasks

Google has launched Gemini 3.5 Flash, now available through the Gemini app, AI Mode in Search, developer tools and enterprise products.

The release introduces a new model family designed for complex agent-based workflows and coding tasks. Gemini 3.5 Pro is also in development and is already being used internally.

Gemini 3.5 Flash becomes the default model for the Gemini app and AI Mode in Search globally. It is also generally available through Google Antigravity, the Gemini API in Google AI Studio and Android Studio, Gemini Enterprise Agent Platform and Gemini Enterprise.

The model is intended to handle long-running tasks that require planning, iteration and execution across several steps. Google positioned it as a system for developing applications, maintaining codebases and assisting with financial document preparation.

Benchmark results published by Google showed Gemini 3.5 Flash outperforming Gemini 3.1 Pro on several coding and agent-based tests, with scores of 76.2% on Terminal-Bench 2.1, 1656 Elo on GDPval-AA, 83.6% on MCP Atlas and 84.2% on CharXiv Reasoning for multimodal understanding.

Google also said the model produces output at four times the speed of other frontier models when measured by output tokens per second. That combination of speed and test performance is central to its push to present the model as suitable for large-scale automated workflows.

Agent focus

Much of the announcement centred on what Google described as agentic use cases, in which software systems carry out multi-step tasks with some level of autonomy under user supervision. Gemini 3.5 Flash can be paired with an updated Antigravity harness to deploy collaborative subagents for more demanding workloads.

Google said this setup can execute multi-step workflows and coding tasks while maintaining performance over longer sequences of actions. It can also generate more interactive web user interfaces and graphics, building on the multimodal features introduced in the Gemini 3 series.

Google linked those functions to workplace automation, saying banks and financial technology groups are using such systems to shorten workflows that previously took weeks. Data science teams are also applying the model to analyse complicated data environments.

Consumer rollout

Beyond developers and business users, Google is deploying Gemini 3.5 Flash in consumer-facing services. The model now underpins Gemini Spark, described as a personal AI agent that helps users manage digital tasks on an ongoing basis.

According to Google, Gemini Spark operates continuously and can take actions on a user's behalf while remaining under the user's direction. The service is beginning with trusted testers before a wider beta release for certain subscribers in the United States.

The same coding and automation features are also being applied across Search, where they will support information agents and more dynamic generative user interfaces. That points to a broader strategy of embedding agent-style AI functions across consumer products rather than confining them to standalone chat tools.

Safety measures

Google said Gemini 3.5 was developed under its Frontier Safety Framework and that it had strengthened cyber and chemical, biological, radiological and nuclear safeguards for the model.

According to Google, those changes are intended to reduce the likelihood of harmful outputs while also cutting mistaken refusals to answer safe prompts. The company said it used new safety training methods and mitigation systems, including interpretability tools designed to inspect the model's internal reasoning before a response is produced.

The launch comes as large technology groups compete to show that AI models can move beyond answering prompts to carrying out structured work across coding, research and administrative processes. In that contest, speed, cost and reliability have become as important as raw benchmark scores, especially for companies trying to turn AI systems into everyday business tools.

Google described Gemini 3.5 Flash as its strongest model yet for agent-based work and coding, intended to help users complete complex tasks in a fraction of the time previously required.