Gemini 3.1 Pro Preview: Access, Context and ARC-AGI-2 Result

One model preview reached several products

Google introduced Gemini 3.1 Pro as a preview. Developers could use it through the Gemini API and AI Studio, as well as tools including the Gemini CLI, Antigravity and Android Studio. Enterprise access ran through Vertex AI and Gemini Enterprise. In the Gemini app and NotebookLM, access was tied to Google AI Pro and Ultra plans. [1]

Atlas interpretation: A launch across many surfaces does not mean each surface offers the same controls. An API developer chooses model settings inside an application; a consumer reaches a configured model through a plan and product interface. Calling both availability is accurate, but the integration work and degree of control are different. [1]

The model card lists a one-million-token input limit and a 64,000-token output limit. These are capacity boundaries, not evidence that every product accepted the same maximum or that a long document would be reasoned over without errors. [2]

The 77.1 percent result needs its test condition

Google reported 77.1 percent on ARC-AGI-2 with Gemini 3.1 Pro set to Thinking High. The launch compared that with 31.1 percent for Gemini 3 Pro under its Thinking High condition. The setting belongs with the number because additional reasoning effort affects both the result and the cost of obtaining it. [1][2]

ARC-AGI presents unfamiliar visual transformation puzzles and asks a system to infer the rule from a few examples. It is useful because memorized subject knowledge is less helpful than discovering the pattern. It does not directly measure source research, factual reliability, software maintenance or workplace completion rates. [3]

Atlas interpretation: The gain supports a narrow conclusion: this model and reasoning setting performed much better on that abstract task suite. Whether the improvement transfers to a product depends on the task, the tools available and whether the answer can be checked. [1][3]

This was an update inside the Gemini 3 line

The November Gemini 3 launch had already distributed the model family across Google's consumer, developer and enterprise products. The 3.1 release upgraded the Pro branch rather than introducing that distribution strategy. [1]

Google's more specialized Deep Think update had arrived a week earlier for research and engineering work. Reading the two events together shows two ways Google exposed more computation: a separately named reasoning mode for selected work, and a new general Pro preview evaluated at a high-thinking setting. [1]