Google and Gemini in 2025: models, products and business

The idea that organises Gemini's year is giving the model more time to think. Deep Think reached the app in August, for now only for Google AI Ultra subscribers, and works by generating several ideas in parallel, evaluating them and combining them before answering. In September, Gemini 2.5 with that capability solved 10 of the 12 problems at the ICPC world programming finals and earned a gold medal; only four of the 139 human teams matched that result.

In November came Gemini 3, which Google presented as the sum of the first generation's multimodality and long context with the second's reasoning and tool use, alongside the Antigravity agent platform. Gemini 3 Flash followed soon after: faster, built into the app and into Search, and in some benchmarks ahead of Gemini 3 Pro itself.

DeepMind also opened two fronts beyond chat. Genie 3 is a world model that generates interactive environments in real time; game developers were sceptical, and DeepMind frames it mainly as a research tool for training agents. Gemini Robotics applies the same logic of reasoning before acting to robots, with one model that turns vision and text into actions and another that plans the steps of a complex task.

Gemini 2.5 Flash Image, known during testing as nano banana, launched in late August with an enterprise pitch: edits that preserve the subject's identity and stay consistent across images. In October Google extended it to Search, Google Photos and NotebookLM, so editing an image with a simple instruction stopped being an isolated feature.

The same image model improved NotebookLM's Video Overviews, which gained six new visual styles, and Google Vids added generative tools based on Veo 3 to turn images into short clips, open to paid Workspace subscribers and to Google AI Pro and Ultra plans.

In December the Gemini app added a provenance check: users can upload a video and ask whether it carries the SynthID mark, an imperceptible signal embedded in audio and imagery, and see which segments are affected. The tool focuses on content created with Google's AI, so it can confirm an origin but not rule one out.

Search's AI Mode went from the US, UK and India to 180 countries and began handling errands, such as finding a restaurant table by group size, date and cuisine. In parallel, Google AI Plus, the more affordable subscription plan that started in Indonesia, opened in 40 more countries, with higher limits for image and video in the Gemini app.

At home and in the car the move was to replace Google Assistant. Gemini took over smart displays and speakers, cameras and doorbells, with a redesigned Google Home app, and in November it began rolling out in Android Auto, first to beta users, with Gemini Live for smoother conversations. The change was not free: contact nicknames, for one, were lost.

For developers, Antigravity drew so much demand that Google reserved higher usage limits and quotas refreshed every five hours for Google AI Pro and Ultra subscribers, and widened the weekly limit for the free plan.

Google Cloud said it is already monetising AI: its backlog reached $106 billion and is growing faster than revenue, with more than half expected to convert into revenue within two years. The routes are usage-based pricing, subscriptions and selling higher-end models and tiers to customers who end up exceeding their commitments.

On the regulatory front, Perplexity offered $34.5 billion for Chrome, far above its own valuation of roughly $14 billion, exploiting the fact that the antitrust remedies under discussion included a possible separation of the browser. The bid was above all a signal of how much capital is chasing distribution.

Risks grew at the same pace. Researchers showed that hidden instructions, for example in a calendar invitation, could push Gemini into controlling Google Home devices; Google said it had applied fixes. And Android's new developer verification system, with free and paid tiers, may affect alternative stores such as F-Droid.