Gemini 3 promises more quality and precision than ever in its responses. The question is whether we will really notice the difference

Google has announced the launch of Gemini 3its new artificial intelligence model. in the company They claim it is their most advanced reasoning model because it is “designed to understand depth and nuance.”

Gemini 3 will also be available as standard as part of AI Mode in the renewed Google search engine (in this case and for the moment, only in the US). It is the first time that Google offers the benefits of its AI model from day one in the search engine, but it also reaches the Gemini app and the developers who work with AI Studio and Vertex AI.

Behind him success of Gemini 2.5 Pro and Flashthe new version arrives in 30 new languages, including Catalan, Basque and Galicianand as we say you can start testing today in the United States… or outside of there via a VPN.

Gemini 3 promises. At least in the tests

Google highlights how the model’s behavior has been outstanding in various synthetic tests. Thus, Gemini 3 leads the LMArena classification with 1,501 points—the first to overcome the 1,500-point barrier.

In fact, he manages to reason “at the level of a PhD” according to the tests of Humanity’s Last Exam (exceeds 37.5% of the test without tools) and GPQA Diamond (91.9%). It also makes spectacular progress in mathematics, as demonstrated by the 23.4% on the MathArena Apex test: GPT 5.1 scores 1.0% and Claude Sonnet 4.5 1.6% on the same test, for example.

The model also wants to be more direct: his answers are more “concise (…) and he prefers to offer valuable information instead of resorting to clichés and flattery. Tells you what you need to hear, not just what you want to hear“.

Gemini 3’s ‘Deep Think’ mode goes even further in tests: in Humanity’s Last Exam it achieves 41.0%, but it also in the demanding ARC-AGI 2 It achieves 45.1% (with code execution), which also demonstrates progress in abstract reasoning and visual understanding.

Gemini 3 explains the world to you in a simple way

The model has a context window of up to one million tokens, which allows it to be used, for example, to analyze huge repositories of code or text and then work on that data.

Its multimodal support allows you to analyze all types of information. For example, Gemini 3 can decipher and translate handwritten recipes in different languages ​​to create a family cookbook that you can share.

Or analyze your pickleball games (we assume the same thing happens in other sports) and identify areas where you can improve and generate a training plan. Or scrutinize the data from a research paper and from it generate code for an interactive guide that helps us better understand those studies.

In fact, integration with Google Search is an especially important part of Gemini 3, which being “embedded” in AI Mode It has the capacity to generate interactive visual elements (widgets, calculators, simulations) in real time. At Google they want the search to be more interactive than ever, and that will mean that sometimes the answers will not be just text, but rather a small interactive webapp that allows us to better understand the answer.

Programming (and agents) to power

The other crucial element of the model is its capacity in the area of ​​programming. Its results in tests of this type are once again outstanding, and for example it tops the WebDev Arena leaderboard with a score of 1,487 ELO.

It also scores 54.2% on Terminal-Bench 2.0, which evaluates a model’s ability to use tools and operate a computer through a terminal. Additionally, it far outperforms 2.5 Pro in SWE-bench Verified (76.2%), a benchmark test that measures the effectiveness of scheduling agents.

These Gemini 3 programming capabilities are intended to be used in a new agent development platform called Google Antigravity. The developer experience is using a “conventional” AI integrated development environment (IDE), but your agents can have access to the editor, terminal, and browser.

That means these agents can autonomously plan and execute complex software tasks and validate their own code, making it easier for human developers to review and audit that code than ever before.

The real challenge of the most recent models

On paper Gemini 3 is postulated as a model that can really make a difference compared to its competitors. The test results and Gemini’s own trajectory make us think that the behavior of this model will indeed be remarkable.

However The question is whether we will really notice the difference. In recent months we have seen how other AI companies have launched new models, but the impact for a large majority of users has been discreet: the previous models already performed really well, and although the new ones undoubtedly provide improvements, for many consultations these improvements allow us to perceive that jump in performance.

Here we see two ways for Google to effectively demonstrate the capabilities of these models. The first opportunity for Gemini 3 will likely be in the area of ​​programming, and it will be these professionals who will likely be able to get the most out of those additional capabilities.

But for the rest of the users, it will be that new AI Mode and the Gemini app that will have to make us notice those features. We are intrigued by this ability to respond with small interactive elements —graphics, widgets—, and perhaps with them we will really discover this new capacity of this chatbot.

In Xataka | Let’s say goodbye to Google Assistant a decade later. Google has begun to delete its code to leave only one option: Gemini