Gemini 3.1 Pro: Google's new model handles logic far better

What's new

Google released Gemini 3.1 Pro on February 19, 2026. It is an update to the company's large AI model, which Google describes as built for the most complex tasks.

The model is already available in many places. In the Gemini app, it is rolling out with higher limits for users with Google AI Pro and Ultra subscriptions, and in NotebookLM only Pro and Ultra users can get it. Developers and businesses can use it as a preview (an early version) via the Gemini API, AI Studio, Antigravity, Vertex AI, Gemini Enterprise, Gemini CLI and Android Studio. It is also available in Microsoft's tools such as GitHub Copilot, Visual Studio and Visual Studio Code. The sources do not make clear whether everything is open in Denmark and the EU.

What's smart

The big number is ARC-AGI-2, a test of whether an AI can solve entirely new logic patterns it has not seen before. Here, Gemini 3.1 Pro got a verified score of 77.1 percent, according to Google. That is more than double Gemini 3 Pro, which scored 31.1 percent. Even Gemini 3 Deep Think, a version aimed at hard problems, reached only 45.1 percent.

In practice, this means the model should be better at tasks where you have to reason your way forward rather than simply reproduce something familiar.

Cheaper or better?

On the API side (the access developers and businesses use to build AI into their own solutions), Gemini 3.1 Pro Preview costs $2 per million input tokens and $12 per million output tokens at the standard tier for prompts up to 200,000 tokens. Tokens are the chunks of text the model reads and writes. Above 200,000 tokens, the price rises to $4 for input and $18 for output.

The Batch and Flex tiers cost 50 percent less, while Priority costs 1.8 times the standard price. Context caching, where the model stores a long text for reuse, costs $0.20 per million tokens ($0.40 above 200,000) plus $4.50 per million tokens per hour in storage. Grounding with Google Search provides 5,000 free queries per month and then costs $14 per 1,000. According to Google, content from paying API users is not used to improve Google's products.

For regular users, the monthly price of Google AI Pro and Ultra is not stated in the sources. Competitors' prices are not documented here either.

Against the competition, Gemini 3.1 Pro generally scores higher than Anthropic's Opus 4.6 and OpenAI's GPT-5.2 on the benchmarks Google highlights, according to The Register. However, Opus 4.6 is ahead on Humanity's Last Exam and SWE-Bench Verified, among others. The figures come from Google's own presentation, and benchmarks should be taken with a grain of salt, because AI models are hard to measure and compare.

What it's good for

The model is aimed at complex tasks where logic and problem-solving carry great weight. Developers can try it directly in tools they already use, such as GitHub Copilot and Visual Studio Code. Businesses can test it via Vertex AI and Gemini Enterprise.

Because it is a preview version, pricing and terms may change. It is therefore smartest to test it on specific tasks before building fixed workflows around it.