Google releases Gemini 2.5 Pro, its first "thinking model"

Background

Most large language models are trained to produce an answer directly, token by token, in a single pass. A thinking model works differently: before giving its final answer it generates intermediate reasoning steps, spending more computation at the moment a question is asked. That extra computation helps with problems that cannot be solved by recall alone, such as multi-step mathematics, science questions, code and puzzles; the trade-off is speed and cost against accuracy.

The announcement

On 25 March 2025, Koray Kavukcuoglu announced Gemini 2.5 and introduced its first member, Gemini 2.5 Pro Experimental ^[1]. Google described Gemini 2.5 as a thinking model designed to tackle increasingly complex problems ^[1]. The Experimental label marked the release as a preview rather than a finished product, and the model was made available to Gemini Advanced subscribers and to developers in Google AI Studio ^[1].

Key details

Google highlighted two measurements at launch. First, the model debuted at #1 on LMArena by a significant margin ^[1]. LMArena is a public leaderboard where anonymous models are compared in blind side-by-side votes cast by users, so a top placement reflects how people prefer one model's answers over another's in open comparisons rather than performance on a single lab-run test.

Second, it scored 18.8% on Humanity's Last Exam among models without tool use ^[1]. Humanity's Last Exam is a benchmark assembled from difficult expert-written questions spanning many disciplines, built so that a high score requires reasoning rather than simple recall. The qualifier matters: the 18.8% figure applies to models answering on their own, without calling search engines, code interpreters or other external tools.

The release itself was staged. The preview went to paid subscribers and to developers through Google AI Studio at the end of March, and Gemini 2.5 Pro became generally available on 17 June 2025 ^[1][2]. In announcing the wider release, Google presented it as an update to its family of thinking models, treating reasoning-first behaviour as a line of models rather than a single experiment ^[2].

Why it matters

The naming was as significant as the numbers. By calling Gemini 2.5 a thinking model, Google put inference-time computation, the work a model does while answering and not only while training, at the centre of how it described its flagship system ^[1]. For users of Gemini Advanced and for developers building on Google AI Studio, the practical effect was access to a model that reasons through a problem step by step before responding ^[1].

The launch also set a pattern that later Gemini 2.5 releases followed: a preview under an experimental label, measurement on public leaderboards and hard benchmarks alongside the announcement, and then a move to general availability a few months later ^[1][2].