Google unveils Gemini 4 models and new AI safety tools

New models for complex tasks

Google recently released Gemini 4 Argon, a frontier reasoning model designed to handle intricate tasks. The company made the model available globally after conducting initial safety testing. It is intended for use in professional workflows such as financial analysis, legal work, and complex coding.

The model is accessible through several platforms, including Google Cloud, the Gemini API, and Google AI Studio. Beyond office tasks, Google is also utilizing Gemini 4 Argon for robotics to power physical-world reasoning. This allows the AI to understand and interact with physical environments more effectively.

In addition to Argon, Google announced Gemini 4 Carbon as its strongest model for long-horizon engineering. This model is currently undergoing pre-release testing with a subset of trusted cyber defenders and the U.S. government. Google is refining safety measures before a broader public launch for this cybersecurity-focused tool.

Simulations and creative tools

The company introduced Genie 4, which functions as a real-time interactive world simulation tool. It features advanced physics, lifelike human movement, and native audio capabilities. Users can interact with these scenes using natural language commands as the environment unfolds.

Google stated that Genie 4 could be used to create safe training environments for AI agents and robots. By simulating complex physics, the tool allows for safer testing of autonomous systems. This helps developers prepare machines for real-world interactions without the risks of physical accidents.

For creators, Google launched Playground, an experimental platform for building and sharing games from text prompts. This tool allows users to create single- and multiplayer games without needing to write any code. Additionally, music producers can use Google Flow Music to generate custom instruments and audio plugins.

Safety and scam detection

Google is expanding the SynthID Detector globally to help users identify AI-generated media. The tool can check if images, videos, or audio clips were created using artificial intelligence. The company has already watermarked more than 180 billion images and videos, as well as 240,000 years of audio.

The verification tools are being built into Search, Chrome, and the Gemini app to handle over a million requests daily. SynthID Detector is capable of identifying AI content from several major tech companies, including OpenAI, NVIDIA, Kakao, and Apple. This provides a layer of verification for content across the web.

On the mobile side, Android is adding an opt-in Enhanced protection mode to Google Messages. This mode uses Private AI Compute and Gemini models to detect sophisticated conversational scams. Google noted that its existing scam detection tools already flag more than 5 million suspected scam calls and texts every month.

Health and research advances

Google is integrating AI more deeply into health technology with the new Google Fitbit Edge. This fitness tracker features precision GPS and automatic recognition for over 50 different activities. It also includes a Gemini-powered Google Health Coach to provide personalized guidance to users.

The company also rolled out updates for Google Health and Fitbit Air, including features for monitoring blood pressure and insulin resistance. In a study published in The Lancet, researchers evaluated the AMIE conversational medical AI system in primary care clinics. The system helped clinicians prepare for urgent care appointments in 75% of cases and matched doctor diagnoses 90% of the time.

In the realm of pure science, Google Quantum AI demonstrated a new method for "magic state cultivation" on its Willow quantum processor. This technique reduces errors by roughly 40-fold compared to uncultivated states. Producing high-purity magic states previously threatened to use up to 95% of a quantum computer's resources, making this a significant step toward fault-tolerant computing.

Space and open models

Google achieved a milestone in space with a prototype Project Suncatcher satellite. The satellite returned the first AI response generated by a data center-class Google TPU in orbit. The response was a haiku composed by the Gemma 4 model, proving the hardware could survive a launch and run workloads in space.

The team plans to continue testing these chips under extreme thermal and radiation conditions. This research helps determine how AI hardware can function in harsh environments outside of Earth. It also showcases the portability of Google's specialized AI hardware.

For the developer community, Google introduced EmbeddingGemma 2, an open embedding model based on the Gemma 4 architecture. It has 740 million parameters and can map various types of data into a shared space on everyday devices. Its modular design reduces local vector storage by up to 6x, which supports private and offline search applications.