Gemini Gets Music: DeepMind Signs U.S. Government Deal

Google DeepMind quietly launched more than 10 AI products and partnerships in four months — and most flew under the radar. The standout reveal: Gemini can now create music, while DeepMind simultaneously embedded itself into the U.S. Department of Energy and the UK government's AI security framework.

This matters because it signals a fundamental shift in strategy. While OpenAI and Anthropic compete for chat users and developers, DeepMind is taking a different path: becoming the AI infrastructure of governments and the creative layer underneath tools millions already use every day.

Gemini AI: Music, Video, and Interactive World Generation

In February 2026, Google DeepMind announced that Gemini gained music creation capabilities — giving the model an expressive creative dimension it previously lacked. This puts Gemini in direct competition with Suno (a text-to-music service where you type a prompt and get a full song back) and Udio (a competitor known for high-fidelity audio output favored by independent musicians).

The same month brought Gemini Deep Think — a dedicated reasoning module for hard mathematical and scientific problems. Think of it as a structured problem-solving engine layered on top of Gemini's standard abilities: not a chatbot response, but a step-by-step logical process designed for researchers, engineers, and STEM students.

On the video side, Veo 3.1 arrived in January 2026 with improvements across three dimensions:

• Consistency — characters and objects stay coherent across frames, solving a major flaw in AI video (where faces often morph unpredictably mid-clip)
• Creativity — more varied visual outputs, moving beyond the flat "stock footage" aesthetic of earlier AI video tools
• Control — users can steer motion, camera angles, and visual style more precisely than before

Veo 3.1's main competitor is OpenAI's Sora (a text-to-video model that went viral in 2024) and Runway ML (a professional video AI used by filmmakers). Fixing cross-frame consistency — one of Sora's early weaknesses — gives Veo a meaningful edge for anyone producing branded content, short films, or educational video.

Project Genie: AI That Builds Playable Worlds

Also in January 2026, DeepMind unveiled Project Genie — a system that generates interactive digital environments (playable 2D game levels, interactive simulations, and explorable virtual scenes) from a text prompt or image. Unlike traditional game engines that require programmers to write code for every rule and physics interaction, Genie learns the logic of a world and generates it dynamically.

This places DeepMind in territory previously owned by Unity and Unreal Engine — but from an entirely AI-native direction. For non-technical creators — marketers building interactive demos, teachers designing learning simulations, indie game developers prototyping ideas — this could compress weeks of development into minutes. The technology is still in research phase, but it's already generating serious attention inside the game design community.

The Government Strategy: U.S. and UK at the Same Time

The most strategically significant development of this period is DeepMind's simultaneous government partnerships, all announced in December 2025. In roughly 30 days, DeepMind signed 3 separate government agreements:

• Genesis Partnership (U.S. Department of Energy) — A national mission to accelerate scientific innovation. The DOE oversees 17 national laboratories including Lawrence Berkeley, Argonne, and Oak Ridge — covering climate modeling, nuclear physics, and drug discovery.
• UK AI Security Institute Partnership — A formal agreement to strengthen AI safety frameworks with the UK's dedicated AI oversight body, which reports directly to government ministers.
• UK Government Partnership — A broader collaboration supporting "prosperity and security in the AI era" — language that typically signals infrastructure-level advising and early access to policy formation.

What the Genesis Partnership Actually Does

Embedding DeepMind's AI tools — likely AlphaFold-descended (protein and molecular structure prediction technology) systems and scientific reasoning models — into DOE national laboratories could accelerate discoveries that would otherwise take decades. For reference: AlphaFold (DeepMind's protein-folding AI, released publicly in 2022) already cut years of biology research down to hours. Genesis scales that ambition to the full breadth of U.S. government science.

For context: OpenAI has separately pursued military and defense contracts with the Pentagon. DeepMind's approach — science and safety institutions first, commercial applications second — is a deliberate contrast that positions it as the "responsible government AI partner." Whether that framing holds over time remains to be seen, but it's a coherent strategic bet that has now won agreements in 2 of the world's most influential governments.

The Honesty Problem — and the Benchmark DeepMind Built to Measure It

The most revealing announcement from this period may be the FACTS Benchmark Suite (December 2025) — a systematic tool for measuring how factually accurate large language models (AI systems trained on massive text datasets to generate human-like responses) actually are in practice.

The fact that DeepMind built and published this benchmark implies an uncomfortable acknowledgment: current AI models, including Gemini, have a hallucination problem (where an AI confidently states false information as established fact). The FACTS Suite lets researchers and institutions:

• Test models across diverse factual domains — science, history, geography, medicine, and current events
• Compare factuality rates between different AI systems objectively
• Track improvement across model versions over time
• Identify subject areas where a specific model is weakest

This matters for everyday users in a concrete way. A teacher using Gemini to check historical facts, a doctor querying drug interactions, a journalist verifying a statistic — all are directly affected by hallucination rates. The FACTS Benchmark is DeepMind's attempt to turn "trust but verify" into an engineering metric instead of a vague disclaimer.

Alongside Gemma Scope 2 (an open-source, freely available tool for AI safety researchers to examine exactly how language models store and represent information internally), it signals that DeepMind is betting transparency about AI limitations — not just performance benchmarks — will become a genuine competitive advantage in the years ahead.

10 Initiatives in 90 Days: What It Means for You

Stepping back, the pace of DeepMind's announcements from December 2025 through February 2026 is striking. Here's the full scorecard:

• January 2026: Project Genie (interactive world generation), Veo 3.1 (video AI upgrade)
• February 2026: Gemini music creation, Gemini Deep Think (math and science reasoning), India AI acceleration program (deploying AI tools for education and scientific advancement)
• December 2025: Gemma Scope 2, Genesis/DOE partnership, UK AI Security Institute partnership, UK government partnership, FACTS Benchmark Suite

That's 10 distinct initiatives across 90 days — roughly 1 major announcement every 9 days. The breadth spans creative tools (music, video, interactive worlds), scientific research (Deep Think, Genesis), safety infrastructure (FACTS, Gemma Scope 2), and government partnerships across 2 countries. No single competitor is attempting to cover this many dimensions simultaneously.

If you're a creative professional, educator, researcher, or anyone who uses Gemini regularly — these updates are coming to you whether or not you follow AI news closely. Gemini's music and video creation tools will surface inside Google's existing products over the coming months. Start by exploring how to integrate AI automation into your daily workflow, and follow our AI news feed to know exactly when each feature becomes publicly available in your region.

Related Content — Get Started | Guides | More News