Google releases Gemini 3.1 Pro in preview

Google announced Gemini 3.1 Pro on Feb. 19, 2026, calling it an upgrade to the model’s core intelligence.[1] Google Cloud describes it as a more capable baseline for complex problem-solving, while Google says it is designed for tasks where a simple answer is insufficient.[2][1] Google DeepMind says the model can comprehend text, audio, images, video and entire code repositories.[3] It lists a context window of up to 1 million tokens and outputs of up to 64,000 tokens.[3] Those limits may be relevant when teams assess long or multimodal inputs.

Developers can access Gemini 3.1 Pro in preview through the Gemini API, Google AI Studio, Gemini CLI, Google Antigravity and Android Studio.[1] Google Cloud says it is also available in preview through Vertex AI and Gemini Enterprise.[2] Consumers can access it through the Gemini app and NotebookLM.[1] Teams should confirm that the specific product and access path they plan to use are included.

Google says Gemini 3.1 Pro scored 77.1% on ARC-AGI-2, more than double Gemini 3 Pro’s reasoning performance.[1] Artificial Analysis ranked the preview first on its Intelligence Index, four points ahead of Claude Opus 4.6, and said it led six of the index’s 10 evaluations.[4] The same evaluator ranked it first on its multimodal MMMU-Pro benchmark and its Coding Index.[4] However, Artificial Analysis reported a score of 1,316 that remained behind Claude Sonnet 4.6, Opus 4.6, GPT-5.2 (xhigh) and GLM-5.[4] These results describe specific evaluations, not every use case.

Artificial Analysis says Gemini 3.1 Pro Preview cost less than half as much as frontier models from OpenAI and Anthropic to run the Intelligence Index.[4] It reports pricing of $2 per million input tokens and $12 per million output tokens for contexts up to 200,000 tokens.[4] That pricing detail and the evaluator’s cost comparison are useful starting points for a model review, but the comparison is tied to a particular benchmark run.[4] Other reported results include a fall in the AA-Omniscience hallucination rate from 88% to 50%; the remaining rate is a reason to assess output quality for the intended task.[4]