Google has officially introduced Gemini 3.1 Pro, the latest update in its Gemini 3 family of models.
The company describes the release as the upgraded core intelligence behind last week’s Gemini 3 Deep Think update and says it is now bringing those gains to a wider audience. The release marks Google’s attempt to reclaim the top position in the fiercely competitive AI race, a throne it has traded back and forth with OpenAI and Anthropic in recent months.
In a blog post announcing the model, Google said: “3.1 Pro represents a step forward in core reasoning.” The company added that the model is “a smarter, more capable baseline for complex problem-solving.”
Built for problems that need more than a quick answer
Gemini 3.1 Pro focuses on deeper reasoning rather than simple responses. Google says the model is intended for situations “where a simple answer isn’t enough,” including research, engineering workflows, data synthesis, and advanced creative work.
The biggest improvement shows up in logic performance.
On the ARC-AGI-2 benchmark, a test designed to measure how well AI handles unfamiliar reasoning patterns, the model scored 77.1%, more than double the performance of Gemini 3 Pro. Google confirmed that this reflects a significant leap in the system’s ability to work through new problems step by step.
The model also posted strong results on other metrics: 94.3% on GPQA Diamond for scientific knowledge and 80.6% on SWE-Bench Verified for coding tasks. Independent evaluations from Artificial Analysis already rank it as the current leader of the pack.
The company says that reasoning boost translates into practical outputs. Examples shared by Google include generating animated SVGs from text prompts, configuring a live aerospace dashboard that visualizes the International Space Station’s orbit, and building immersive interactive experiences, such as a 3D starling murmuration controlled with hand tracking.
Google also highlighted creative coding tasks, including designing a modern website inspired by Emily Brontë’s Wuthering Heights, in which the model interprets tone and theme rather than simply summarizing the text.
According to the Google DeepMind model card, Gemini 3.1 Pro is a “natively multimodal reasoning” system that can process text, images, audio, video, and large codebases. It supports a context window of up to 1 million tokens and produces outputs of up to 64,000 tokens, enabling analysis of large datasets or lengthy workflows in a single session.
Wait, is it better at everything?
Not quite. Benchmark data show a few areas where the new version actually slipped slightly. On MMMU Pro, a multimodal understanding test, Gemini 3 Pro still edges out the newcomer 81.0% to 80.5%. And on Humanity’s Last Exam with tool support, Anthropic’s Opus 4.6 took the top spot.
Safety evaluations also showed minor fluctuations. According to DeepMind’s documentation, image-to-text safety dipped 0.33% compared to the previous version, though the company noted that manual review confirmed most flagged cases were either false positives or not actually problematic. Unjustified refusals, those moments when a model wrongly declines to answer, improved slightly by 0.08%.
“We continue to improve our internal evaluations, including refining automated evaluations to reduce false positives and negatives,” the DeepMind team noted in their documentation.
Who gets it now
The rollout is broader than usual for a point release.
Consumers can access 3.1 Pro through the Gemini app starting today, while NotebookLM users with Pro and Ultra plans get access to it too. For developers, it’s available in preview through Google AI Studio, the Gemini API, and even inside Android Studio. Enterprise customers can find it in Vertex AI and Gemini Enterprise.
Google is calling this a “preview” for now, so they can gather user feedback before making it the official standard.
For more on Google’s AI ambitions, check out how Gemini Lyria 3 is pushing into AI-powered music generation.