Google Gemini 3 Pro: The End of the Chatbot Era, the Beginning of the AI Agent Age?

Gemini 3 Pro is presented as Google’s most «reasonable» model, and the company openly positions its solution as the industry leader in many respects. Independent evaluations confirm this.

According to Artificial Analysis, the model has become the new leader of their integral index.

If Artificial Analysis’ tests are to be believed, Google has taken the lead over its competitors in the areas of intelligent tasks — reasoning, understanding complex structures, accuracy and multimodality.

The performance in Deep Analysis deserves special attention. On Humanity’s Last Exam, which assesses a model’s ability to solve doctoral-level problems without tools, Gemini 3 Pro scored over 37%.

This is more than ten percentage points higher than the previous record. On ARC-AGI-2, one of the most challenging benchmarks that assesses the ability to derive rules and apply them to new situations, the model also scored above most competitors.

The high performance is also evident in the math tests, Google stressed. In the MathArena Apex test, where questions of extreme levels of complexity traditionally take models out of balance, Gemini 3 Pro received 23.4%. Previously, this figure was unattainable for other systems, and the best results did not exceed 5.2%.

In multimodal tests, the updated Gemini also takes the first positions. Experts directly attribute this to the potentially large scale of the model.

This hypothesis would explain the ability of Google’s AI to outperform products from other companies in tasks involving visual analysis and spatial understanding.

Separately, a comparison with Claude and ChatGPT is worth noting. On the SWE-Bench Verified benchmark, which tests the ability to autonomously handle GitHub tasks, the new model lags behind Sonnet 4.5 by only one percent. In other metrics, the Gemini often comes out ahead.

Another important piece of evidence is the speed of the model. Artificial Analysis notes that Gemini 3 Pro generates about 128 tokens per second. This is faster than the performance of GPT-5.1, Kimi K2 Thinking, and Grok 4.

This is most likely due to Google’s own hardware platform based on Tensor Processing Unit (TPU) processors.

Thus, in a number of parameters, the model confidently competes with existing flagships and in many cases surpasses them. At the same time, the product lags behind its competitors in some tests, but usually only slightly.

Technically speaking, Gemini 3 Pro is a multimodal model with more context and an expanded set of controllable parameters. According to Google’s documentation, it accepts text, code, images, audio, video, and PDFs as input.

The maximum input size is claimed to be 1,048,576 tokens, with output up to 65,536 tokens. For practical applications, this means that the model can analyze a large amount of data at once, including long documents, sets of articles, large video lectures or entire code repositories.

The documentation separately states that the model supports up to 900 images per query, up to 900 documents, up to ten videos and audio lasting up to several hours.

This makes it possible to build complex queries where text descriptions are combined with visual material and code.

With the introduction of Gemini 3, numerous new settings are introduced. Firstly, a reasoning level parameter. Instead of the previous «thinking budget», an explicit thinking_level switch is used, which can be low or high.

Google explains this as a way to adjust the amount of internal reasoning. The feature strikes a balance between response quality, logic complexity, latency, and cost.

Secondly, resolution control for media content has appeared. The media_resolution parameter (low, medium, or high) allows controlling the depth of visual analysis and token consumption for images and videos.

At the product level, Google is trying to turn these features into new forms of interfaces. Pichai describes the Gemini experience this way:

In addition, the Google executive writes that Gemini 3 «brings powerful reasoning to search and new generative interfaces». Specifically, it introduces a visual layout mode.

This means that an answer can look like a «magazine» spread with photos, modules, and items that the user controls. As an example, he cited a request to plan a three-day trip to Rome.

The system responded by generating an itinerary with visual blocks and the ability to customize it to the user’s preferences.

An important area of development is agent-based capabilities. Google notes that since Gemini 2, the company has been actively developing the «agent era» and Gemini 3 shows progress in the ability to plan actions over long stretches of time.

Their materials specifically mention leadership on the Vending-Bench 2 benchmark, which simulates the management of a vending machine over the course of a year. According to Google’s description, the AI keeps the sequence of actions and the use of tools in a stable state and does not «move away» from the set goal.

The company attributes the practical application of these ideas for everyday tasks to the Gemini Agent. A Google executive explains that this feature uses advanced reasoning capabilities to break down complex tasks into multiple steps.

After completing them, the model suggests further actions to the user, depending on the result obtained.

On the engineering side, the new Google Antigravity development tool has become an important element of the ecosystem. The official announcement describes it as an «agent-based development platform».

The solution is an integrated environment where Gemini 3-based agents have access to an editor, terminal, and browser. They can plan and execute complex software tasks and present their steps to the user as individual «artifacts» that are easy to inspect.

The ArsTechnica publication emphases that Antigravity can use not only Gemini-based agents, but also Claude Sonnet 4.5 and GPT-based solutions. The product also offers client and server-side command-line tools.

These cases show that the model can handle tasks related to spatialisation of objects and working in virtual and augmented reality environments.

In summary, at the feature level, Gemini 3 Pro looks like a versatile tool with great context, guided deep reasoning, and tight coupling with development tools and agents.