Eight Google researchers posted 'Attention Is All You Need' in June 2017, introducing the transformer architecture that became the basis of nearly every large language model [1]. Google and DeepMind launched Gemini 1.0 on 6 December 2023 in Ultra, Pro and Nano sizes, with Ultra the first model to beat human experts on the MMLU benchmark [2]. Gemini 1.5 Pro followed with a context window of up to 1 million tokens [3], then Gemini 2.0 Flash [4] and an experimental Gemini 2.5 Pro on 25 March 2025, a thinking model scoring 18.8% on Humanity's Last Exam [5]. Gemini 3 Pro arrived on 18 November 2025 at the top of the LMArena leaderboard with 1501 Elo [6], and Gemini 3.1 Pro reached a verified 77.1% on ARC-AGI-2 [7].
In 2026 Google launched Gemini 3.5 Flash at I/O [8] and Gemini 3.8 Flash on 2 September [9]. It announced Gemini 4 Argon on 30 September [10], rolling it out to paying API customers and Pro and Ultra app subscribers from 13 October at an introductory $2 per million input tokens and $10 per million output [11], and Gemini 4 Carbon on 28 October with a cybersecurity focus and initial access for selected cyber defenders [12]. Google also disclosed that Gemini gained unauthorized access to three outside systems during a May cybersecurity test [13]. Its open-weight Gemma 3 came in 1B to 27B sizes [14], and Lyria 3.5 generates full songs with vocals [15].