Google Introduces Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber

As organizations move from experimenting with artificial intelligence to deploying it across real business processes, the requirements for AI models are changing. Intelligence alone is no longer enough. Production AI agents must also respond quickly, operate reliably and complete tasks without consuming unnecessary computing resources.

Google is addressing these requirements with three new additions to the Gemini model family: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and Gemini 3.5 Flash Cyber. Announced on July 21, 2026, the new models are designed to make AI agents faster, more capable and more cost-effective at scale.

Rather than offering a single model for every scenario, the latest Gemini lineup gives developers and enterprises more specialized options for complex reasoning, high-volume automation and cybersecurity.

Gemini 3.6 Flash: The New Workhorse for AI Agents

Gemini 3.6 Flash is positioned as Google’s new workhorse model for coding, knowledge-based work and multimodal tasks. It builds on Gemini 3.5 Flash while improving model quality, token efficiency and the ability to complete multi-step workflows.

According to Google, Gemini 3.6 Flash uses 17% fewer output tokens than Gemini 3.5 Flash on the Artificial Analysis Index. In certain coding evaluations, Google observed token reductions of up to 65%. The model also requires fewer reasoning steps and tool calls to complete multi-stage tasks.

This matters because the true cost of an AI agent is not determined only by the published price of individual tokens. An agent may need to reason, call external tools, review results and repeat actions before completing a task. Reducing unnecessary steps can lower the total cost and latency of the complete workflow.

Gemini 3.6 Flash is priced at $1.50 per one million input tokens and $7.50 per one million output tokens. Google says this lower pricing, combined with improved efficiency, reduces the overall cost of building and operating agentic systems.

The model also demonstrates improvements across coding, computer use and professional knowledge work. Google reports stronger results in evaluations covering software engineering, machine learning research and real-world computer interaction. It has also been used for document parsing, chart analysis, financial-data processing, report drafting and code migration workflows.

For enterprises, these capabilities could support agents that analyse reports, work across business applications, assist developers and complete structured knowledge tasks with less manual intervention.

Gemini 3.5 Flash-Lite: Built for Speed and Scale

Not every AI task requires the most advanced model. Many enterprise workflows involve large numbers of relatively focused operations, such as classifying documents, extracting information, processing search results, translating content or summarising transactions.

Gemini 3.5 Flash-Lite is designed for these high-volume, low-latency workloads. Google describes it as the fastest and most cost-effective model in the Gemini 3.5 series. It can generate approximately 350 output tokens per second, according to the Artificial Analysis Index.

The model is priced at $0.30 per one million input tokens and $2.50 per one million output tokens. It also introduces configurable thinking levels, allowing developers to balance speed, cost and reasoning depth depending on the complexity of each request.

For straightforward, repetitive workloads, developers can use minimal or low thinking levels to prioritise faster execution. More complex sub-agent tasks can use higher thinking levels when additional reasoning is required.

Google highlights potential applications such as agentic search, document processing, extracting features from large ecommerce datasets, translating and summarising receipts, generating design concepts and rapidly testing multiple development options.

This makes Flash-Lite particularly relevant for organizations expecting AI agents to handle thousands or millions of routine actions. A lightweight model can process these tasks while a more capable model, such as Gemini 3.6 Flash, coordinates the broader workflow.

Gemini 3.5 Flash Cyber: Specialized AI for Software Security

The third model, Gemini 3.5 Flash Cyber, focuses specifically on cybersecurity. Built on Gemini 3.5 Flash, it has been fine-tuned to find, validate and help fix software vulnerabilities.

The model operates within CodeMender, Google’s code-security agent. Multiple Gemini 3.5 Flash Cyber agents can work together and combine their findings into a single security report. Google says the model delivers competitive results on CyberGym, a benchmark used to evaluate vulnerability discovery and repair capabilities.

Because cybersecurity models can have both defensive and potentially harmful applications, Google is taking a controlled approach to availability. Gemini 3.5 Flash Cyber will initially be offered through CodeMender to governments and trusted partners as part of a limited-access pilot.

The objective is to help defenders identify and resolve critical software weaknesses before attackers can exploit them, while limiting the risk of broader misuse.

A More Specialized Approach to Enterprise AI

The introduction of these three models reflects a broader shift in how production AI systems are being designed. Instead of sending every request to one large general-purpose model, organizations can assign different models to different responsibilities.

Gemini 3.6 Flash can manage complex reasoning, coding and multimodal analysis. Gemini 3.5 Flash-Lite can execute high-volume tasks quickly and economically. Gemini 3.5 Flash Cyber can support specialized defensive security operations.

Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are available through the Gemini API, Google AI Studio and Android Studio. They are also accessible through Google’s enterprise Gemini platforms and the Gemini app, while Flash-Lite is beginning to roll out in Google Search.

For businesses planning their next generation of AI agents, the message is clear: successful enterprise AI will depend not only on choosing the most powerful model, but on selecting the right combination of intelligence, speed, specialization and cost for every workflow.