The AI ​​arena begins to clear. Just now, Google pulled the trigger again and officially launched Gemini 3 Flash. This is another violent output after the Gemini 3 Pro. Without any warning or any foreshadowing, Google directly announced that Gemini 3 Flash has now become the default model in Gemini applications, completely replacing 2.5 Flash. This means that hundreds of millions of users around the world can immediately experience the inference capabilities of the Gemini 3 series models without paying any fees.


If Gemini 3 Pro is to take full advantage of AI computing power, then Gemini 3 Flash breaks the impossible triangle between "high intelligence", "low cost" and "fast response".

Opening the Model Card, we see a surprising set of data: Gemini 3 Flash scored as high as 78% on SWE-bench Verified, the authoritative benchmark that evaluates the capabilities of coding agents. This not only leaves the previous 2.5 series far behind, but even surpasses its big brother Gemini 3 Pro in some areas, such as logical depth. What's even more outrageous is that while providing this "crushing" performance, it costs less than a quarter of the Gemini 3 Pro.

This may not only be a victory for the Waiting Party in terms of cost performance, but more like an unreasonable "muscle show" by Google.

Relatively speaking, Gemini 3 Flash is more suitable for some development work scenarios that require high frequency and extreme speed. With extremely low latency, Gemini 3 Flash can update applications at almost real-time speed. Different from waiting for a long time for response in the past, Gemini 3 Flash response can become a "brain" that quickly completes reasoning, error correction, and self-verification in a large-scale complex flow.

For ordinary users, Google has thrown out another "king bomb": zero-threshold voice website building. This means that you don’t need to know any code, you just need to describe your ideas casually to Gemini, and Gemini 3 Flash can transform those scattered ideas into a fully functional application in a few minutes.

Although Gemini 3 could achieve this to a certain extent before, with Gemini 3 Flash, the price is lower, the workflow is simpler, and the time cost is lower. Currently, Gemini 3 Flash is priced at $0.50 per million input tokens and $3 per million output tokens, with audio input still priced at $1 per million input tokens.

From video analysis, data extraction to visual question and answer, Gemini 3 Flash, coupled with the iteration of search algorithms, is also redefining the response limits of AI. It is currently available simultaneously through Google AI Studio, Gemini API and Vertex AI. This wave of "fast, accurate and ruthless" releases by Google has announced that in the large model arena, the last barrier to speed and intelligence has been demolished. The new king has arrived and is everywhere.


Gemini 3 Flash launched in Google AI Studio ｜Source: Geek Park

01

This time, "lightweight" no longer means "compromise"

The core value of the release of Gemini 3 Flash is not just a simple parameter change, but that small models can also surpass some flagship models in the core capabilities of Agent. In the SWE-bench and Toolathlon tests that measure agent coding and long-range tool invocation, Gemini 3 Flash not only outperformed its big brother Gemini 3 Pro, but even suppressed the top models of GPT and Claude in certain dimensions.

It can also be seen that in automated work scenarios that require frequent interaction and rapid feedback, shorter reasoning links and higher sensitivity to follow instructions may have more practical value than huge parameter scales.


Gemini 3 Flash demonstrates ultra-high intelligence in various top benchmark tests | Source: Google official website

Of course, this does not necessarily mean that models with large parameters have no application value. Although Gemini 3 Flash has achieved nearly 7 times improvement compared to 2.5 Pro on visual reasoning puzzles such as ARC-AGI-2, there is still a certain gap between it and the top SOTA model when dealing with extremely complex architectural designs. This also means that the positioning of Gemini 3 Flash is not all-powerful, but partial enhancement.

But more importantly, Gemini 3 Flash provides a lower entry barrier for the upcoming era of intelligent agents by lowering the input cost to $0.50 and combining it with substantial caching discounts, and also creates conditions for an explosion. You know, maybe a year ago, it was very expensive to get this kind of PhD-level reasoning ability, but now it's almost free. It can also be seen that large models still cannot escape price wars under the competition of homogeneous technology. Obviously, Google currently has the advantage in this game.

In terms of specific performance, according to third-party analysis benchmarks, Gemini 3 Flash runs a full three times faster than 2.5 Pro. Logic evolution and extremely low latency make it accurate and fast when processing tedious tasks such as high-volume legal contracts and extracting definition terms.


Gemini 3 Flash breaks through the Pareto limit in terms of performance, cost and speed | Source: Google official website

In the multi-modal field, Gemini 3 Flash has demonstrated remarkable dominance in video understanding and complex chart analysis, proving that Google's internal "perception is reasoning" capability has matured. In particular, it can transform complex unstructured video data into executable business plans in seconds, which means that visual information is no longer the special expertise of AI, but part of the underlying logic. Perhaps the vast amounts of dormant data on Google Chrome can once again be activated as liquid business assets.

For developers and enterprise users, Gemini 3 Flash directly lowers the threshold for cutting-edge AI deployment to a freezing point through extremely competitive pricing and contextual caching technology. Whether it is supporting online customer service conversations or realizing automatic programming of agents through Google Antigravity, it is proving that: high performance, low latency and extremely low cost can be achieved at the same time by choosing Gemini 3 Flash now.

Today, the Flash series of models is no longer an "alternative" that exists for compromise, but has become a weapon more suitable for mass developers to upgrade. The arrival of Gemini 3 Flash may promote the large-scale explosion of intelligent agents to a certain extent and accelerate the arrival of the intelligent agent application era.

02

Violent upgrade of search efficiency:

Google Search’s Last Piece of the Model Puzzle

Starting from the second half of this year, search has clearly become the focus of Google. Gemini 3 Flash is also online and is directly sent to the search system. To a certain extent, we can also see that the current model upgrade is no longer just an upgrade of a single product line, but a coordinated improvement of the entire AI product ecosystem.

First, Gemini 3 Flash will be rolled out globally and directly become the default configuration of Google search AI mode. As long as users use Google AI to search, they will directly feel the power of the Gemini 3 series models.

The mutual exclusion between deep reasoning capabilities and instant response speed is no longer an eternal problem for models. Gemini 3 Flash's improvements in reasoning capabilities, tool calling, and multi-modal processing allow the system to produce more structured and logical responses when responding to detailed inquiries under complex constraints without sacrificing the crucial timeliness in search scenarios. This also means that in the past"Higher-order reasoning" is transforming into a standardized infrastructure for mass retrieval, and AI search can also move from simple information matching to real-time answers to complex questions.

At the same time, for higher task requirements, the introduction of Gemini 3 Pro and Nano Banana Pro into the search field has also filled the gap in the vertical field to a certain extent.

Combined with the "Thinking with 3 Pro" model currently launched by Google in the US market, it can be seen that Google is not trying to create conventional AI retrieval, but hopes to be able to perform dynamic visual layout and interactive simulation presentation of heavy computing tasks such as complex mathematical programming. With the addition of Gemini 3 Flash, Google has implemented a relatively comprehensive model product layout based on user needs: Flash is responsible for high-frequency, extremely fast inclusive intelligent interactions, and Pro is responsible for low-frequency but high-value logic tasks. Obviously,Future AI interaction will definitely not be a single model fighting alone, but a dynamic allocation of computing power and intelligent layering based on task complexity.

The emergence of Gemini 3 Flash objectively marks the shrinking of the "intellectual gap" between small models and large models.It proves that after algorithm optimization reaches a certain threshold, the bottleneck of intelligent experience is no longer the scale of computing power, but how to use this extremely fast intelligencefeelSeamlessly woven into the user's daily decision-making flow.With the parallel provision of "Quick Mode" and "Thinking Mode", AI interaction has officially evolved from "experimental dialogue" to an industrial-grade assisted decision-making engine. As for the model family bucket as the technical base, Google has already prepared it for everyone.

03

After the model left the laboratory, the Google ecosystem once again broadened its boundaries

Just now, the balance of the AI ​​model ecosystem has tilted again. The emergence of Gemini 3 Flash and the full rollout of Google's Gemini 3 series models mean that the ecological advantages of Google's models have been strengthened again, and are triggering chain reactions in the task cycles of various vertical industries.

In the field of software engineering, coding platforms such as Cursor and Devin have found that the intervention of Gemini 3 Flash allows the AI ​​response speed to keep up with the engineer's intuition, allowing the "coding agent" to change from an asynchronous waiting process to a near-real-time synchronous collaboration.

In scenarios such as law and finance, which have almost stringent requirements for accuracy, the practice of Harvey and Box AI has proven that Gemini 3 Flash can achieve a 15% improvement in accuracy on tasks such as complex financial data identification and cross-referencing of long contracts without sacrificing speed. This also shows that AI is finally able to process high-volume unstructured data at an industrial level, and no longer requires users to make a painful choice between "deep understanding" and "real-time feedback."

In addition, deepfake detection platform Resemble AI leverages its multimodal capabilities to instantly transform complex forensic data into concise intelligence, analyzing it 4x faster than before, while Bridgewater uses it to capture those fleeting conceptual understandings in large-scale multimodal data sets.

Even in the field of game development, Latitude uses its near real-time inference performance to move the character logic of the game world from preset scripts to true autonomous intelligence.


Image source: Google official website

As can be seen, Gemini 3 FlashSuccessfully running through the last mile from prototype development to large-scale implementation proves that the best technology should not only be the advantage of a few people, but should be the cornerstone of an era that welcomes a large-scale explosion of productivity.