Google Releases Three Gemini Flash Models: AI Competition Enters the Cost Era

Google Drops Three New Gemini Flash Models Late at Night

Late last night, Google released three new Gemini Flash series models in one go, covering three distinct product lines: performance, security, and cost. This update is more than a routine version iteration—it reveals a noteworthy industry signal: the competitive focus in AI large models is quietly shifting from "who's smarter" to "who's cheaper."

According to analysis from relevant content creators on Bilibili, these three models target different application scenarios, forming a complete product matrix. Google also revealed that its flagship model Gemini 4 has launched "the largest-ever pre-training run," meaning an even higher performance ceiling is on the way.

Positioning Differences Among the Three Gemini Flash Models

Before diving into the specific models, it's important to understand where the Flash series sits within the Gemini product matrix. Gemini models are tiered by capability, typically including Ultra (flagship), Pro (professional), and Flash (lightweight high-speed). The Flash series' design philosophy is to maximize inference speed and cost efficiency while maintaining sufficient intelligence—not every task requires maximum compute power, and a lighter model can handle the majority of everyday tasks.

Gemini 3.6 Flash is the strongest performer in this release, with its most critical highlight being a 17% reduction in token consumption compared to the previous generation. Tokens are the basic unit of measurement for how large language models process text—each Chinese character corresponds to roughly 1-2 tokens, and cloud providers typically charge based on the number of input and output tokens. A 17% reduction in token consumption means the model generates fewer redundant tokens when producing equivalent-quality responses, or employs more efficient attention mechanisms internally (such as sparse attention or KV Cache optimization), thereby reducing the actual compute overhead per API call. At a time when LLM inference costs remain high, this optimization directly impacts enterprise usage costs.

Gemini 3.5 Flash Cyber is called the world's first AI model dedicated to cybersecurity. Cybersecurity has always been an important direction for AI applications, but general-purpose large models have obvious limitations when handling security tasks: they lack deep understanding of specialized knowledge such as CVE vulnerability databases, attack chains, and malicious code signatures, and may pose data leakage risks when processing sensitive security data. The Cyber model likely incorporates large volumes of security logs, threat intelligence, and vulnerability reports during the pre-training phase, with targeted optimization for security analysis scenarios during the alignment phase. This is a highly vertical product positioning, indicating that Google is beginning to do deep customization for specific industry scenarios rather than trying to cover everything with a single general-purpose model.

Gemini 3.5 Flash Lite focuses on being the fastest and cheapest, purpose-built for high-concurrency AI Agent scenarios.

Here it's worth explaining why Agent scenarios are so cost-sensitive. AI Agents are AI systems that can autonomously plan and execute multi-step tasks, fundamentally different from traditional single-turn dialogue models. A typical Agent might need 5-20 rounds of internal inference calls to complete a single task, including task decomposition, tool calling, and result verification. When thousands of Agents run simultaneously, the backend model faces concurrent request volumes that can be dozens of times higher than ordinary chat scenarios. Even tiny cost differences multiply dramatically with call volume, creating enormous billing gaps—this is precisely the core pain point Flash Lite aims to solve.

From the positioning of these three models, we can see that Google is splitting the Flash series into increasingly fine segments: choose 3.6 for performance, Cyber for security, and Lite for running Agents. This kind of granular product strategy is itself a sign of market competition reaching maturity.

AI Competition Shifts from "Smarter" to "Cheaper"

The most intriguing aspect of this release is that Google promoted "17% cost reduction" as one of its core selling points.

Keep in mind that Google, as a top-tier player in AI, has traditionally attracted attention with model capabilities and benchmark scores. When even Google starts emphasizing cost, it actually reflects a transformation in the competitive logic of the entire industry.

Why Large Model Cost Has Become So Important

As large model capabilities gradually converge, the performance gap between top models is narrowing. Once baseline capabilities are "good enough," what enterprise customers truly care about becomes:

• For the same task, which model runs cheaper?
• In high-concurrency scenarios, can costs be pushed to the minimum?
• After large-scale deployment, are long-term inference costs controllable?

For enterprises that need to run AI 24/7, a 17% reduction in token consumption isn't an abstract number—it's a tangible difference on the bill. Taking a mid-sized enterprise consuming 100 million tokens per month as an example, at current mainstream API pricing, a 17% savings could mean a cost difference of thousands to tens of thousands of yuan per month. When AI applications scale up to millions of calls, such cost optimizations are amplified dramatically.

This also explains why Google specifically launched a "cheapest" Flash Lite—AI Agent high-concurrency scenarios are extremely cost-sensitive. An intelligent agent that requires multiple rounds of calls and continuous operation often has its commercial viability determined by cost control.

How Falling Costs Are Changing Enterprise AI Applications

As models compete to become cheaper, the most direct beneficiaries are enterprises looking to deploy AI at scale.

Take AI customer acquisition as an example—a high-frequency, high-concurrency application scenario. These applications typically require models to run around the clock handling massive interactions, making them extremely cost-sensitive. In the past, high inference costs were a major barrier preventing many small and medium enterprises from scaling up AI adoption.

From Tools to "AI Employees"

As cost floors continue to drop, the form of AI applications is also evolving. A batch of intelligent agent products has already emerged that make AI customer acquisition "real"—they're no longer just simple conversation tools, but take on complete workflows approaching that of an "employee":

• 24/7 automatic monitoring of short video comment sections, precisely identifying potential customers with genuine purchase intent;
• Conducting one-on-one in-depth engagement using compliant, professional messaging;
• Connecting Douyin, WeChat, and WeCom across three platforms to build a private traffic ecosystem.

From lead generation to conversion, then to repeat purchases and referral growth, the entire pipeline can be automated by intelligent agents, dramatically reducing manual intervention. The reason these applications can achieve viable business logic is precisely because the cost of underlying large models keeps declining—the cheaper the large model, the lower the marginal cost of customer acquisition for enterprises. This is why every price reduction by Google, OpenAI, and other major players directly stimulates innovation bursts in the downstream application layer.

Industry Signals Worth Watching

Overall, Google's simultaneous release of three Gemini Flash models conveys several trends worth watching.

First, product line granularity. Google is no longer covering all needs with a single model, but segmenting by performance, security, and cost, indicating that enterprise customer needs have become sufficiently diverse and mature.

Second, cost as a core competitive advantage. When top players all treat cost reduction as a selling point, it means the arms race in model capabilities is giving way to competition over efficiency and value for money.

Third, deep cultivation of vertical scenarios. The Cyber model dedicated to cybersecurity foreshadows more industry-specific models in the future, with general-purpose and vertical models developing in parallel.

Meanwhile, Google revealed that Gemini 4 has launched the largest pre-training run in history. Pre-training is the foundation of large language model capabilities, referring to the process of unsupervised learning on massive text data. "Largest-ever scale" typically means more training data (potentially reaching tens of trillions of tokens), larger model parameter counts, and more compute cluster investment. Google possesses its own proprietary TPU chips and world-class data center infrastructure, giving it unique advantages in pre-training scale. This shows that the pursuit of "stronger" hasn't stopped—it's just that while chasing greater strength, "cheaper" has become an equally important pole. For enterprises, this is undoubtedly welcome news as dividends continue to be released.

Key Takeaways