Gemini 3.1 Pro 'Most Powerful AI Ever'—Google Strikes Back!
The new model scores 77.1% on ARC-AGI-2, more than doubling its predecessor's reasoning capability.
Google has fired back in the AI arms race with Gemini 3.1 Pro, a refined version of its Gemini 3 series that retakes the title of most powerful AI model. The headline achievement is a 77.1% score on the ARC-AGI-2 benchmark, which tests an AI's ability to solve novel logic patterns—more than doubling the performance of its predecessor. This release follows a pattern where Google, OpenAI, and Anthropic have been leapfrogging each other every few weeks, with Google briefly holding the top spot last year before being overtaken. Crucially, Google is offering this massive reasoning upgrade at no additional cost, maintaining the same API pricing of $2 per 1M input tokens and $12 per 1M output tokens.
Beyond raw reasoning, Gemini 3.1 Pro excels in specialized areas, scoring 94.3% on GPQA Diamond (PhD-level science questions) and 92.6% on the multimodal MMMLU benchmark. Google also released a separate API endpoint (gemini-3.1-pro-preview-customtools) for developers building with custom tools and bash, improving the model's focus on user-specified tools. The model is immediately available in Google AI Studio, Vertex AI, the Gemini CLI, and GitHub Copilot for Pro/Business users, with consumer rollout underway in the Gemini app. This release signals that models available today for logic, synthesis, and multi-step planning are categorically different from those six months ago, enabling more sophisticated AI-powered features.
- Doubled reasoning capability with 77.1% on ARC-AGI-2 benchmark, testing novel logic pattern solving
- Maintained competitive pricing at $2/$12 per 1M tokens despite massive performance gains
- Immediate availability across developer platforms (AI Studio, Vertex AI, GitHub Copilot) and consumer app rollout
Why It Matters
Developers get a free, significant upgrade for building advanced AI features involving logic and multi-step planning.