Gemini 3.1 Pro

What it is

Gemini 3.1 Pro is a large language model from Google, released in preview on 19 February 2026 as an upgrade to Gemini 3 Pro. Google described it as "a step forward in core reasoning", and said the model was the upgraded core intelligence behind the Gemini 3 Deep Think update of the previous week.^[1] The same reasoning improvements therefore arrived first inside a deep-thinking release and then as a general Pro model.

Benchmark result

The headline figure Google reported comes from ARC-AGI-2, a benchmark built from puzzle-like tasks where memorised answers help little and a system has to infer a rule from a handful of examples before applying it. Gemini 3.1 Pro scored 77.1% on it in a verified run, which Google said was more than double Gemini 3 Pro's reasoning performance.^[1]

A single score says only so much, but it points at a specific ability: working out an answer to a problem the model has not seen, rather than retrieving one. For tasks that require several steps of derivation, that ability matters more than recall.

Where it was available

In preview, developers could reach the model through the Gemini API in Google AI Studio, Gemini CLI, Google Antigravity and Android Studio, and enterprises through Vertex AI and Gemini Enterprise.^[1] Consumers got it in the Gemini app, with higher limits on the Google AI Pro and Ultra plans, and in NotebookLM for Pro and Ultra users.^[1][4]

Pricing

API pricing for the preview was $2 per million input tokens and $12 per million output tokens for prompts up to 200,000 tokens, and $4 and $18 per million above that.^[2] Batch API pricing was $1 per million input tokens and $6 per million output tokens for prompts up to 200,000 tokens.^[2] Context caching cost $0.20 per million tokens for prompts up to that size, plus a storage fee of $4.50 per million tokens per hour.^[2] Google's pricing page stated that the preview was not available on the free tier.^[2]

The shape of these numbers is easy to read. A request that carries more context gives the model more tokens to process, so the per-token rate rises once a prompt passes 200,000 tokens. Batch work is cheaper because requests can be grouped and processed together, and caching is cheaper again because a long context that is reused does not have to be recomputed — though the storage fee means a cache only pays off if it is actually read from.

Place in the lineup

Gemini 3.1 Pro replaced gemini-3-pro-preview, a model identifier that was shut down on 9 March 2026, so applications pointing at the old name had to move to the new one.^[3] As of October 2026 it was the Gemini Pro model offered in Google Antigravity.^[5]

The release pattern is itself worth noting: a preview launch, a dated shutdown for the model it replaced, and per-token prices published at the same time. For anyone deciding whether to build on the model, those three details are what make migration cost and running cost calculable in advance.