字母榜精选

wrote a column ·

The smallest iteration of Google's Gemini yet has prompted Yao Shunyu to exclaim 'unstoppable'—why?

People are still talking about the amusing scene where the bosses of OpenAI and Anthropic refused to shake hands, leaving them with clenched fists raised high. Meanwhile, Google quietly pushed out a new model iteration.

This iteration seems like a wolf in sheep's clothing—judging by the numeric suffix in the model name, this is Google’s smallest iteration yet. Previously, updates jumped from Gemini 2.0 to Gemini 2.5, but this time it’s from Gemini 3.0 to Gemini 3.1 Pro Preview.

However, this '.1' iteration has not improved by just a small margin.

Google CEO Sundar Pichai stated that the new model excels at handling 'super complex tasks,' such as visualizing complex concepts, synthesizing data into a single view, or bringing creative projects to life.

Yao Shunyu also posted on X to promote Gemini 3.1 Pro Preview, praising it enthusiastically:

“Gemini is not just a good model; an even better model is arriving with unstoppable momentum.”

Notably, about a week ago, Google launched 'Dedicated Reasoning Mode' Gemini 3 Deep Think, designed for complex, open-ended problems in science, research, and engineering.

Gemini 3 Deep Think was the first project Yao Shunyu worked on after moving from Anthropic to Google DeepMind.

The current Gemini 3.1 Pro Preview has intricate ties to Gemini 3 Deep Think. Officially, it is said to be 'directly built upon the experience and technology of Gemini 3 Deep Think,' essentially bringing the core reasoning enhancements of Deep Think down to the more widely accessible Pro model.

What can Gemini 3.1 Pro Preview do?

Since the standout capability of this new generation model lies in handling 'super complex' tasks, ordinary conversations are set aside. In Google's official blog post, several examples were highlighted to showcase its capabilities.

First, create SVG animations with simple prompts.

This feature was available in the previous generation, but the improvement is notably significant by comparison.

For example, the prompt 'Generate an SVG depicting a chameleon quietly sitting on a tree branch. Make the chameleon’s eyes follow the user’s mouse cursor as it moves across the screen.'

The animation background generated by Gemini 3 Pro is a plain white, and the chameleon appears rather stiff, even having two eyes on one side.

In contrast, the animation generated by Gemini 3.1 Pro features a rich 'dark green jungle' background. The chameleon’s body is decorated with yellow stripes and dots, its eyes are three-dimensional, and its legs are naturally bent.

Another example is the prompt 'Generate an SVG for a sliding toggle switch. When the mouse hovers over the sun icon, transform it into a glowing moon, while smoothly transitioning the background from bright to dark. Use a clean flat UI style.'

Although the animation provided by Gemini 3 Pro completes the task and the icons change with the mouse, the main icon is simplistic—a circular pattern with a missing segment, using yellow for daytime and white for nighttime.

The animation generated by Gemini 3.1 Pro is much more complex, showing a bright yellow sun with white clouds during the day and a crescent moon with stars at night, with both sets of icons transitioning smoothly.

In summary, the animations created by Gemini 3 Pro remind people of the old meme about 'studying animation for three years.'

The SVG animations delivered by Gemini 3.1 Pro have reached a level where they can be used directly.

Second, construct an engineering-grade system.

Gemini 3.1 Pro can now generate a complete interactive system integrating 3D rendering, real-time solar ephemeris calculation, asynchronous API data fetching, and physical lighting effects based on a highly complex natural language instruction, rather than a simple page demo.

In the example provided by Google, users gave text instructions, and Gemini 3.1 Pro generated a high-fidelity, interactive 3D International Space Station (ISS) orbit tracker using high-resolution Blue Marble texture maps to render a detailed 3D Earth model.

Third, generate interactive creative systems.

In another example, Google showcased a complex 3D murmuration simulation written by Gemini 3.1 Pro.

It not only generates visual code but also builds an immersive experience where users can manipulate the flock of birds through hand tracking and listen to generative soundtracks that change with the movement of the flock.

For researchers and designers, this provides a powerful way to prototype interfaces rich in sensory experience.

Fourth, transform literary themes into executable code.

This example might be the one where the general public can most easily appreciate its impressive capabilities.

When tasked with creating a modern personal portfolio website for Emily Brontë's 'Wuthering Heights,' the model did not merely summarize the text. Instead, it inferred based on the novel's atmosphere and emotions, designing a clean, contemporary interface that captures the spirit of the protagonist.

The value of abstract reasoning ability goes without saying.

How strong is it?

It’s inevitable for new-generation models to undergo benchmark testing.

The '.1' upgrade achieved scores that doubled or even tripled previous results.

According to test results released in Google’s official blog post,

In the ARC-AGI-2 benchmark, version 3.1 Pro scored 77.1% on the validation test, more than doubling the reasoning performance of version 3 Pro.

This aligns with the examples of 3.1 Pro, as this test evaluates the model’s ability to solve entirely new logical patterns—in simpler terms, its abstract reasoning and puzzle-solving capability.

Additionally, in the GPQA Diamond (science knowledge test), 3.1 Pro scored 94.3%; in the agent-based benchmark MCP Atlas, it scored 69.2%; and in BrowseComp, which assesses real-world web browsing and information integration abilities, it scored 85.9%.

These scores surpass those of Anthropic's Sonnet 4.6 and Opus 4.6, as well as OpenAI's GPT-5.2 and GPT-5.3-Codex.

Google's Gemini 3.1 Pro has significantly widened the gap in ARC abstract reasoning and BrowseComp search tasks, showing a clear Agent-oriented tendency rather than being purely a knowledge model.

Moreover, third-party evaluation agencies specializing in large model benchmarking and comparative analysis have released relevant test results, highly praising Gemini 3.1 Pro for leading in six out of ten evaluations that constitute the Artificial Analysis Intelligence Index. Compared to Gemini 3 Pro Preview, it shows significant improvements in multiple capabilities, especially in reasoning and knowledge, coding ability, and reducing hallucinations.

Furthermore, Gemini 3.1 Pro Preview maintains high token efficiency.

Running the full Artificial Analysis Intelligence Index requires approximately 57 million tokens (1 million more than Gemini 3 Pro Preview).

This token usage is lower than other cutting-edge models running in maximum reasoning mode, such as Opus 4.6 (max) and GPT-5.2 (xhigh).

Combined with a lower per-token pricing, Gemini 3.1 Pro Preview has a cost advantage among frontier models, with the cost of running the full Intelligence Index being less than half of Opus 4.6 (max), though still about twice that of the leading open-source model GLM-5.

Double the capability at the same price.

Google’s official API pricing indicates that Gemini 3 Pro/3.1 Pro Preview charges based on token usage:

For fewer than 200k tokens, input costs approximately $2 per million tokens, with output priced at $4. For more than 200k tokens, input costs $4 per million tokens, with output priced at $18.

In terms of context caching, depending on the prompt size, each million tokens incurs a charge of $0.20 to $0.40, plus a storage fee of $4.50 per million tokens per hour.

At this price point, it is consistent with Gemini's previous generation 3 Pro overall, but compared to the Anthropic Opus series, it is relatively affordable. Models like Opus have input/output unit prices around $5/$25.

Especially considering its currently outstanding model capabilities, this price appears even more competitive.

Do not forget, what Google released this time is just a 'Preview,' and Google will soon roll out the official version. The '.1' iteration also hints that it’s merely a small demonstration of strength.

Currently, developers can access 3.1 Pro via AI Studio, Gemini API, Gemini CLI, the agent development platform Google Antigravity, and Android Studio; enterprise users can use it on Vertex AI and Gemini Enterprise; general users can utilize it through the Gemini app and NotebookLM, though the latter is limited to Pro and Ultra subscribers.

In various communities, many users have already jumped in to try it out, creating impressive projects just as shown in Google's demos.

Someone used Gemini 3.1 Pro to generate an interactive 3D mechanical-grade car suspension system simulator, including realistic geometry, linkage constraints, real-time steering, and travel calculations. This is equivalent to combining mechanical engineering modeling, physics logic, and 3D visualization into one operational tool, approaching engineering-grade prototyping capabilities.

Another user created a looping animation titled 'Ghost Hunter Walks Through a Haunted House' using 3.1 Pro, exclaiming 'Gemini is no joke.'

All in all, Google has truly unveiled a major move this time.

A minor '.1' update significantly boosted reasoning and coding capabilities while maintaining stable pricing.

The enthusiasm for hand-crafted demos in the community further proves its capability and practicality.

The AI community is becoming increasingly pragmatic. No matter how powerful the model, ultimately it comes down to whether the bill is worth it. Companies are meticulously calculating the return on each token, and developers must also weigh the cost-effectiveness. Google's move not only aims to reclaim the throne but also pushes the competition into a new phase of 'who can manage their resources better'.

Let’s see how Anthropic and OpenAI, who are holding their cards close, along with competitors like xAI, Meta, and Microsoft, will respond next. $Alphabet-C (GOOG.US)$$Alphabet-A (GOOGL.US)$$AI (LIST0535.SH)$$Artificial Intelligence (LIST2136.US)$$Artificial Intelligence (LIST23586.HK)$$Technology (LIST20763.US)$$Technology (LIST20840.HK)$$Star Tech Companies (LIST2518.US)$

105K Views