Google has launched Gemini 3.1 Proan incremental update of its flagship model that comes loaded with surprises. And according to its benchmarks, the model has much more to say than it seems. In abstract reasoning, Google wants to start setting the pace on Anthropic and OpenAI. But their ace in the hole is not just that, because they have something that other startups cannot replicate: their entire ecosystem and how they are integrating AI into it.
What just happened. Just three months after launching Gemini 3 ProGoogle has published Gemini 3.1 Pro. The curious thing is that the jump is much more impressive than it may seem if we only looked at that “.1” in front of it. According to the company, the new model significantly improves the reasoning of the previous one and represents the intelligence base that already fed the Gemini 3 Deep Think update, presented last week.
It is available today in the Gemini app, in NotebookLM (for Pro and Ultra subscribers), in the API through AI Studio, and in enterprise environments through Vertex AI.
Data. In the ARC-AGI-2 benchmark, designed to evaluate the ability to solve completely new logical patterns, without the possibility of having seen them during training, Gemini 3.1 Pro has achieved 77.1%. To put it in context: Gemini 3 Pro stayed at 31.1%, while Claude Sonnet 4.6 marked 58.3% and Opus 4.6 68.8%. That is, Google has not only closed the gap, but has gone over it.
It should be noted that never before has a mid-term review of its models recorded such a pronounced advance in reasoning.
What the numbers say in the rest of the benchmarks. In the comparative table that accompanies the advertisementGemini 3.1 Pro tops the majority of categories evaluated: it obtains the best result in Humanity’s Last Exam without tools (44.4%), it leads in GPQA Diamond with 94.3% in scientific knowledge, and it doubles the previous model in APEX-Agents, the benchmark for long-term tasks. It also excels in MCP Atlas (multistep workflows), BrowseComp (agent search) and MMMLU (multilingual question and answer).
It should be noted that, according to these benchmarks, it is not better in everything: in GDPval-AA Elo, which evaluates tasks in real-world work environments, Claude Sonnet 4.6 surpasses Gemini 3.1 Pro with 1,633 points compared to 1,317. And in SWE-Bench Verified, the programming test with agents, Opus 4.6 scores 80.8% compared to Google’s 80.6%. However, in the global calculation, the balance clearly favors Google’s new model.
In Arena Leaderboard (the classification based on user votes) still places Claude Opus 4.6 ahead in text and code, although here “the sensations” of each user take more prominence when it comes to rating, than anything else.
A clear competitive advantage. The strongest argument in favor of Google does not even have to do with the power of its latest model. The company doesn’t need to convince you to use its AI: it’s already where you are. Search, Gmail, YouTube, Android, Docs, Drive, Google Photos, Maps… Its AI does not depend on you opening a specific application, but is integrated into the ecosystem that millions of people already use daily.
For the rest of the startups (OpenAI, Anthropic…), they need you to use their models in specific environments (ChatGPT, Claude). Google is simply already there. It’s a moat that perhaps not even the best model in the world could sweep right now.
And then there’s the price. Gemini 3.1 Pro comes to users with a subscription to Google AI Plus, Pro and Ultra, although you can also try it on a limited basis in the free plan. It should be noted that it is currently in a preliminary version.
The narrative that Google wants us to have in our heads is that, for a modest price, you have access to that model, plus everything the company offers in its ecosystem, including storage. That, right now, is very difficult to overcome. Additionally, for developers, the API is also offered at a very competitive price. So, from a practical point of view and from the pocket, Google is giving everything so that all its users continue using its ecosystem, with or without the best AI.
The “.1”. The AI race has been at a frenetic pace for months. And the most interesting of all is that Google, which arrived late for the racehas had a hell of a year in which he has structured all the mess he had with his AI. The jump from Gemini 3 to 3.1 in reasoning is greater than what many rivals have achieved between full versions. And it has done so while maintaining the advantage of being the company that controls the most relevant entry points to the Internet. It remains to be seen how they solve monetizing your artificial intelligencebut they have certainly put in the work.
Cover image | Alex Dudar and Google
GIPHY App Key not set. Please check settings