Gemini

Google had invented the transformer, run the largest models in the world and been caught flat-footed by ChatGPT. In April 2023 it merged Google Brain and DeepMind into one organisation under Demis Hassabis, and on 6 December it released the result. Gemini came in three sizes, Ultra, Pro and Nano, was trained on TPU v4 and v5 pods, and was natively multimodal: rather than bolting an image encoder onto a language model, it had been trained from the beginning on interleaved text, code, images, audio and video. Ultra was the first model to beat human experts on the MMLU benchmark, at 90 percent, and matched or exceeded GPT-4 on most others.

The launch also included a demonstration video that had been edited to look like real-time interaction, and the criticism it drew set the tone for a year in which Google's models were judged harshly. Gemini 1.5 in February 2024 introduced a million-token context window, which changed what the models could be asked to read.

Gemini is on this timeline as the moment the three-laboratory frontier settled, OpenAI, Anthropic and Google DeepMind, and as the model family that in 2025 achieved gold-medal performance at the mathematical olympiad and, as Gemini 3, took the lead on most benchmarks.