What Is Google Gemini? Google's Answer to ChatGPT, From Bard Onward

Gemini is Google's generative AI: today it is both the name of a chat app and the name of the family of models inside it. If you want a one-line answer to "what is Gemini", it is Google's response to ChatGPT, which began as Bard in February 2023, swapped its insides for the Gemini models announced on 6 December 2023, and took the Gemini name itself on 8 February 2024. As of October 2026 the newest large model is Gemini 4 Argon.

Bard: a chat service in a hurry

On 6 February 2023 Sundar Pichai announced Bard, a conversational AI service, on Google's blog. It ran on a lightweight version of LaMDA, the dialogue language model Google had unveiled in 2021, and opened first to "trusted testers". That announcement, Microsoft's new Bing the next day, Bard's factual error in a promotional clip and the fall in Alphabet's shares are covered in that article.

Public access began in the United States and the United Kingdom on 21 March 2023. Google called Bard an "early experiment" and took sign-ups for a waitlist. At Google I/O on 10 May it said Bard now ran on PaLM 2, removed the waitlist, opened to more than 180 countries and territories, and added Japanese and Korean.

6 December 2023: models named Gemini

On 6 December 2023 Pichai and Google DeepMind's Demis Hassabis introduced Gemini as "our largest and most capable AI model." Gemini 1.0 came in three sizes: Ultra for highly complex tasks, Pro for a wide range of tasks, and Nano for on-device use. The announcement described it as the first realisation of the vision Google had when it formed Google DeepMind earlier that year (see DeepMind).

Google stressed two things. One was design: Gemini was "natively multimodal", pre-trained from the start on text, images, audio and other modalities together. The other was scores. Google wrote that Gemini Ultra, at 90.0%, was the first model to outperform human experts on MMLU, a test of knowledge and problem-solving across 57 subjects. That was Google's own measurement, and the same post said it had used a new approach to MMLU that let Gemini "think more carefully" before answering hard questions. The models were trained on Google's own TPU v4 and v5e.

From the same day Bard ran on a fine-tuned version of Gemini Pro, which Google called the biggest upgrade since Bard launched, starting in English in more than 170 countries and territories. Nano shipped on the Pixel 8 Pro, powering features such as summaries in the Recorder app.

The edited demo

What drew the most attention was a video called "Hands-on with Gemini". A person sketched a duck, shuffled a ball under cups and made shadow puppets, and Gemini seemed to watch and answer aloud as it happened.

On 7 December TechCrunch ran the headline "Google's best Gemini demo was faked." According to TechCrunch, Bloomberg's Parmy Olson first reported the discrepancy, and Google's own explanation was that it had captured footage, then prompted Gemini with still frames from that footage and with text. The video itself carried a note that latency had been reduced and Gemini's outputs shortened. Oriol Vinyals, Google DeepMind's VP of Research, said the video illustrated what multimodal experiences built with Gemini could look like and was made to inspire developers.

Some of the actual prompts and responses had been published in a related blog post linked from the video's description. TechCrunch acknowledged that, said a Google spokesperson had asked it to change its headline, and kept the word "faked" on the grounds that the video did not reflect reality.

February 2024: a new name and a paused feature

On 8 February 2024 Google renamed Bard to Gemini. It also launched Gemini Advanced, with access to the top Ultra 1.0 model, as part of a new Google One AI Premium plan at $19.99 a month, available in English in more than 150 countries and territories. Mobile apps for Android and iOS began in the US, with Japanese and Korean promised from the following week. From then on the service and the models shared one name.

A failure followed quickly. According to Raghavan's later account, the Gemini app (still called Bard at the time) had launched image generation that could depict people in early February. In late February images depicting historically white-dominated scenes with racially diverse figures spread on social media. According to the Associated Press, Google acknowledged inaccuracies in historical depictions on 21 February, and on 22 February announced on X that it would pause image generation of people.

On 23 February senior vice president Prabhakar Raghavan explained what happened in a post titled "Gemini image generation got it wrong. We'll do better." He gave two causes. Tuning meant to ensure Gemini showed a range of people failed to account for cases that should clearly not show a range; and over time the model became far more cautious than intended, refusing some prompts entirely by wrongly treating very anodyne ones as sensitive. Raghavan added that Gemini, as a creativity and productivity tool, might not always be reliable on current events, evolving news or hot-button topics, and that hallucinations were a known challenge with all LLMs.

1.5, 2.0 and 3

From then on Gemini's story was told in model numbers. With Gemini 1.5 on 15 February 2024, Google said it could consistently process up to a million tokens, pitching it at reading whole long documents and videos. Gemini 2.0, on 11 December 2024, was billed as a model for the "agentic era". In announcing Gemini 3 on 18 November 2025, Google wrote that Gemini 2.5 Pro had topped the LMArena leaderboard for more than six months.

Gemini 3 shipped in Google Search on its first day. In the same announcement Google said the Gemini app had passed 650 million monthly users and Search's AI Overviews had reached 2 billion monthly users — both the company's own figures. Gemini-branded reasoning models produced results in mathematics too: in July 2025 an advanced version of Gemini Deep Think reached the gold-medal threshold at the International Mathematical Olympiad.

Apple, and Gemini 4

On 12 January 2026 Apple and Google announced in a joint statement a multi-year collaboration under which the next generation of Apple Foundation Models would be based on Google's Gemini models and cloud technology. According to the statement, those models would power future Apple Intelligence features, including a more personalised Siri due that year, while Apple Intelligence would continue to run on Apple devices and Private Cloud Compute.

Model development was not all smooth. According to TechCrunch, Gemini 3.5 Pro, which Google had said in May 2026 would roll out the following month, still had not shipped in July, and Google updated its lighter Flash models first. Gemini 4 Argon, announced in September, has a one-million-token output limit and, according to Google, was released first in phases to trusted cyber defenders. On 13 October it began rolling out globally in the Gemini app to Google AI Pro and Ultra subscribers.

Afterwards

Gemini's history is also the story of three years in which Google played catch-up. A rushed Bard, an edited demo and an image feature paused for overcorrection made the first year better known for its stumbles. After that, Google unified the name and became the one supplying models — to Search, to Android, and to Apple's Siri. The wider story is in the AI timeline, and Google's own in the Google page.