Gemini Updates: Lyria 3, Gemini 3.1, Nano Banana 2 & More

The Gemini app received a substantial update this month, expanding its creative and problem-solving capabilities. Google’s latest “Gemini Drop” introduces Lyria 3, a new music generation model, alongside improvements to Gemini 3.1’s reasoning abilities, enhanced image editing with Nano Banana 2 and tools to streamline video creation, and research. These updates, announced February 18th and 20th, 2026, aim to make the Gemini app a more versatile tool for both casual users and professionals.

Crafting Soundscapes with Lyria 3

Perhaps the most immediately accessible new feature is Lyria 3, Google’s most advanced AI music model. Lyria 3 allows users to generate 30-second musical tracks from either text descriptions or image uploads. This isn’t simply about creating background music; the model is capable of producing tracks with vocals, lyrics, and a range of genres. As Google explains in the official Lyria 3 overview, the goal is to allow users to “turn a funny moment, a specific feeling, or even a photo into a custom soundtrack.” The system is designed to be flexible, allowing users to mix genres, moods, and eras to create personalized audio for various purposes, from focused work sessions to shareable content.

Crucially, Google has incorporated safeguards to respect copyright and artistic expression. When a user’s prompt references a specific artist, Lyria 3 will generate music that evokes a similar style or mood, but won’t directly imitate the artist’s work. This approach, detailed in a Google Blog post, aims to balance creative freedom with ethical considerations. All music generated within the Gemini app is also embedded with SynthID, an imperceptible watermark designed to identify AI-generated content, addressing growing concerns about the provenance of digital media.

Gemini 3.1: Enhanced Reasoning and Scientific Applications

Beyond music creation, the update brings significant improvements to Gemini’s core intelligence with the release of Gemini 3.1. This iteration focuses on enhancing the model’s ability to tackle complex problems, particularly in scientific and engineering fields. Gemini 3.1 Pro, designed for demanding workflows, and Deep Think, a specialized reasoning mode, are now available – the latter exclusively for Google AI Ultra subscribers.

The leap in reasoning capability is particularly noteworthy. According to Google, Gemini 3.1 Pro achieved a verified score of 77.1% on the ARC-AGI-2 benchmark, a rigorous test of a model’s ability to solve novel logic problems. This represents more than a doubling of performance compared to its predecessor, Gemini 3 Pro. This improvement translates to practical applications like synthesizing data, building interactive designs (including coding immersive 3D experiences with hand-tracking), and generating website-ready animated SVGs from text prompts. As reported by Case Western Reserve University’s UTech news, the model can even engage in “creative coding,” adapting the style or tone of an image or document to design a functional website interface.

Visual Fidelity and Streamlined Video Creation

The Gemini app’s visual capabilities have also been upgraded with Nano Banana 2, the latest iteration of Google’s image model. Nano Banana 2 allows users to add text to images in any language with improved accuracy and speed. This feature is particularly useful for creating visually engaging content for social media or presentations. Complementing this is the introduction of Veo Templates, which provide pre-designed video styles that users can customize with their own content, simplifying the video creation process.

Accelerating Research with Verified Citations

For researchers and students, the update includes a valuable new feature: direct links to verified scientific citations. Gemini can now identify and provide links to the original scientific papers supporting information presented, eliminating the need to manually search through extensive databases. This feature enhances the reliability of research and promotes access to high-quality data. This is a significant step towards combating the spread of misinformation and ensuring that information is grounded in credible sources.

Limitations and Considerations

While these updates represent significant advancements, it’s important to acknowledge their limitations. The performance of Lyria 3, while impressive, is still constrained by the 30-second track length. The quality of generated music will also depend heavily on the clarity and specificity of the user’s prompts. Similarly, Gemini 3.1’s enhanced reasoning abilities, while demonstrated on benchmarks like ARC-AGI-2, may not translate directly to all real-world scenarios. The model’s performance can vary depending on the complexity and ambiguity of the task. The availability of Deep Think is limited to Google AI Ultra subscribers, creating a tiered access structure.

The utilize of SynthID watermarks is a positive step towards transparency, but the effectiveness of such watermarks in preventing misuse remains an ongoing area of research. As noted by experts in digital forensics, watermarks can be vulnerable to removal or manipulation, requiring continuous development of more robust techniques.

Looking Ahead: Continued Refinement and Expansion

Google’s ongoing development of the Gemini app suggests a commitment to continuous improvement and expansion of its capabilities. Future updates are likely to focus on refining the performance of Lyria 3, expanding the range of genres and styles it can generate, and increasing the length of generated tracks. Further enhancements to Gemini 3.1’s reasoning abilities are also anticipated, potentially leading to even more sophisticated applications in scientific research and engineering. The company will likely continue to address ethical concerns related to AI-generated content, exploring new methods for ensuring transparency and protecting intellectual property. The rollout of these features will continue across countries where the Gemini app is available, with ongoing monitoring and adjustments based on user feedback and performance data.