Google releases Gemini 3.1 Pro

Background

Gemini is the name Google gives its family of generative AI models, and the company makes them available through several channels at once: a consumer chat app, a developer API and cloud platforms aimed at businesses. Benchmarks are how the field compares models, and ARC-AGI-2 is one of the harder ones: it poses abstract reasoning puzzles that are meant to be unfamiliar rather than drawn from text a model may have seen during training, so a score on it is read as a rough signal of how well a system handles problems it has not encountered before.

The release

On 19 February 2026, Google released Gemini 3.1 Pro in preview ^[1]. The announcement put the emphasis on difficult, multi-step work, and the number Google highlighted came from the reasoning side. On ARC-AGI-2 the model scored 77.1%, more than double the score of Gemini 3 Pro ^[1].

What the demos showed

Google's demonstrations were built around the model producing working artefacts rather than text alone. They included animated SVGs and a live dashboard of the International Space Station's orbit ^[1]. Both are outputs where a model has to hold structure and detail together at the same time: the drawing has to remain coherent while it moves, and the dashboard has to keep its layout while its values track something real.

Where it was available

The consumer rollout was in the Gemini app, where users select "Pro" in the model menu to reach the model ^[2]. Google said the app rollout came with higher limits for users on the Google AI Pro and Ultra plans ^[1][2]; usage caps of this kind are the practical constraint on how much of a model a subscriber can actually use.

For developers, Google said Gemini 3.1 Pro could be used in preview through the Gemini API in Google AI Studio ^[1]. Google AI Studio is the company's browser-based environment for trying out and building with its models, and the Gemini API is the interface developers call from their own code. For enterprises, Google said the model could be reached through Vertex AI and Gemini Enterprise ^[1] — respectively its cloud platform for building with models and its offering for organisations.

Status

Google said general availability would follow soon ^[1]. A preview is not the final version of a model: access arrangements can change and the model can be updated before the general release, so the 19 February launch is best read as the point at which the model became publicly usable rather than the point at which it was finished.

Why it matters

Two things stand out in the announcement. The first is the size of the ARC-AGI-2 jump: more than doubling the previous Pro model's result is the kind of gap that gets used to argue a model has changed qualitatively, not just improved at the margin ^[1]. The second is the breadth of the same-day availability. The app, the Gemini API in AI Studio, Vertex AI and Gemini Enterprise were all listed at once ^[1][2], which meant individual users, developers and companies could reach the same model from the first day rather than waiting for a staged rollout by audience.

For anyone tracking how Google positions its models, the release also shows the shape of the company's launch pattern: a preview first, with paid consumer plans getting more headroom, and general availability announced as a later step ^[1][2]. The headline benchmark number is the part that gets quoted, but the distribution — app, API, cloud — is what determines who can actually build on the model.