Gemini 3 Deep Think

What it is

Gemini 3 Deep Think is an enhanced reasoning mode of Google's Gemini 3 model family — a way of running the model that is meant for questions where a fast answer matters less than a carefully worked-out one. Google described it as a mode that pushes Gemini 3 performance even further^[1]. It was announced together with Gemini 3 Pro on 18 November 2025^[1].

A reasoning mode of this kind does not change what the model is; it changes how it answers. Instead of replying immediately, the model is given room to work through a problem at length before it commits to an answer. That makes such modes slower than ordinary prompting, which is why they are offered as an option rather than as the default.

Release and access

Google did not open Deep Think to everyone at once. It first gave the mode to safety testers, then rolled it out to Google AI Ultra subscribers, and released it on 3 December 2025^[1][2]. In Google's apps, users selected the Deep Think option in the prompt bar, and Gemini notified them when their results were ready^[3] — a design that fits a mode whose answer may take noticeably longer than usual.

Reported results

Google published three benchmark figures for Deep Think: 41.0% on Humanity's Last Exam without tools, 93.8% on GPQA Diamond, and 45.1% on ARC-AGI-2 with code execution, verified by ARC Prize^[1]. Humanity's Last Exam is a set of expert-written questions built to be hard for frontier models; GPQA Diamond is a set of graduate-level science questions; ARC-AGI-2 is a puzzle-style test of solving unfamiliar problems. The three numbers therefore measure different things, and the Humanity's Last Exam figure was reported without tool use while the ARC-AGI-2 figure allowed code execution.

For comparison, Google's launch post listed Gemini 3 Pro at 37.5% on Humanity's Last Exam without tools^[1]. The same post gave Gemini 3 Pro's position on the LMArena leaderboard — where people compare model outputs without knowing which model produced them — at 1501 Elo^[1]. Google called the 45.1% ARC-AGI-2 result with code execution unprecedented^[1].

The upgraded version

On 12 February 2026 Google released an upgraded Gemini 3 Deep Think aimed at science, research and engineering problems. It was made available to Google AI Ultra subscribers through a "Deep Think" option, with results usually ready within minutes^[3]. Google later said that Gemini 3.1 Pro, released a week later, was the upgraded core intelligence behind those gains^[4].

Where it sits

Deep Think is best understood as a tier within the Gemini 3 family rather than a separate model beside it: the same underlying intelligence, offered in a mode that trades speed for depth. Announcing a reasoning mode alongside the flagship model, opening it first to the top subscription tier, and then following with an upgraded version pitched at scientific and engineering work — that is the shape its public history has taken so far.