Google launches Gemini 3.1 Pro and promises AI smarter than ever

Just three months after the launch of Gemini 3 Pro, Google is already unveiling a new version of its flagship model.

With Gemini 3.1 Pro, the American giant is changing its pace…and its naming convention. Goodbye to ".5" updates, hello to a ".1" version that reflects the ambition to move faster than its rivals.

Performance that shakes up OpenAI and Anthropic

What if Google regained the lead with its new Gemini 3.1 Pro model?

– Source: Google

Behind this announcement, one figure stands out: 77.1% on ARC-AGI-2, a benchmark dedicated to abstract reasoning. This score puts Google ahead of its main competitors.

In detail, Gemini 3.1 Pro surpasses GPT-5.2 Thinking by 34.5%, and Claude Opus 4.6 by 40% on Humanity’s Last Exam, and by 68.8% on ARC-AGI-2, according to the rankings mentioned. Even on Humanity’s Last Exam, a notoriously demanding academic test, Google boasts a 44.4% score, ahead of competing models. In agentic use cases—scenarios where AI must autonomously utilize tools, code, APIs, and web browsing—the progress is equally significant. On APEX-Agents, Gemini 3.1 Pro achieves 33.5%, compared to 18.4% for the previous version. It also takes the lead on BrowseComp with 85.9% and MCP Atlas with 69.2%, two benchmarks for evaluating the ability to orchestrate complex, multi-step tasks. Finally, another noteworthy technical element is the introduction of the "thinking_level" parameter. It replaces the old "thinking_budget" and allows you to adjust the depth of internal reasoning according to the complexity of the task.

Rapid integration into the Google ecosystem

A preview of the new capabilities of Gemini 3.1 Pro with the same prompt – Source: Google

For this new version, Gemini 3.1 Pro retains a context window of one million tokens and integrates immediately with the entire Google ecosystem. after transforming its AI into a personal assistant.

Google AI Pro and Ultra subscribers can already use it in the Gemini app and NotebookLM, and for developers, the model is available via the Gemini API, AI Studio, Vertex AI, Gemini CLI, and Android Studio.

For other users, Google promises availability in the coming weeks, after incorporating feedback from this testing phase. The strategy remains well-established, where Google first deploys new versions to professionals and advanced users, before adjusting behaviors in real-world conditions to broaden access. With Gemini 3.1 Pro, Google is showing other market players that the battle is only just beginning in advanced AI reasoning, but competitors have undoubtedly not had their last word…