Google wins again. Gemini 3.1 Pro review

Gemini 3.1 Pro: A Detailed Review & Performance Analysis

Key Concepts:

• Gemini 3.1 Pro: Google’s latest and most performant AI model, excelling in multi-modal capabilities (text, image, audio, video).
• Context Window: The amount of information an AI model can process in a single prompt (Gemini 3.1 Pro boasts a 1 million token window).
• Hallucination Rate: The tendency of an AI model to generate factually incorrect or nonsensical information.
• Benchmarks: Standardized tests used to evaluate the performance of AI models (e.g., Humanity’s Last Exam, ARC AGI 2, GPQA).
• Multi-modality: The ability of an AI model to process and understand different types of data (text, images, audio, video).
• Emergent Abilities: Unexpected capabilities that arise in large AI models, such as pattern recognition and learning.

---

I. Introduction & Accessibility

The video introduces Gemini 3.1 Pro as Google’s most powerful AI model to date, highlighting its broad capabilities. Currently, access is available through the Gemini app (selecting the “Pro” option), NotebookLM, Google AI Studio, Gemini CLI, Google’s Anti-gravity IDE, and Android Studio, as well as various enterprise platforms. The presenter emphasizes the immediate availability of the model at the time of recording.

II. Creative Capabilities & Fluid OS Example

A demonstration showcases Gemini 3.1 Pro’s creative potential by prompting it to design a new mobile operating system, “Fluid OS,” superior to Android and iOS. The OS includes eight core apps:

• Omni: A unified AI agent providing proactive information based on calendar, email, and location.
• Thread: A consolidated messaging platform integrating SMS, WhatsApp, email, and DMs.
• Sense: An app aggregating biometric data from wearables and phone sensors.
• Flow: A universal media player.
• Prism: An AR lens for real-time translation, text copying, and visual search.
• Shift: A navigation app with proactive traffic updates.
• Vault: A secure storage for passwords, credit cards, and keys.
• Home: A smart home control center.

The presenter notes the lack of a definitive “correct” answer, inviting audience feedback on the OS design.

III. 3D Generation & Spatial Understanding

Gemini 3.1 Pro demonstrates superior performance in 3D generation compared to other models. A prompt to create a 3D animation from an image of a pagoda resulted in a more detailed and refined output than competing models. Specific improvements noted include enhanced detail, smoother animations, and more accurate representations in SVG animations. The model’s ability to generate detailed 3D models from single prompts is highlighted.

IV. Music Composition & Interface Creation

The video demonstrates Gemini 3.1 Pro’s ability to compose music. First, the model was prompted to create a piano roll interface for note input. Subsequently, a request for a “powerful, expressive 32 bar piano opus” yielded a harmonious and complex composition. This performance was contrasted with JLM5, another state-of-the-art model, which produced a less coherent result. The model exhibits inherent music composition knowledge.

V. Lighting Physics Simulation & Parameter Control

Gemini 3.1 Pro was used to create a simulation featuring three metallic spheres suspended above a street scene. The model successfully rendered realistic reflections between the spheres, and allowed for adjustable parameters like reflectivity, roughness, and color. The presenter highlights the ability to achieve a fully functional simulation with parameter control in just two prompts.

VI. Practical Applications & Use Cases

Several practical applications are showcased:

• Receipt Parsing: The model accurately parsed multiple receipts, extracting data (date, item, total, currency) and exporting it to a Google Sheet, even handling receipts in different currencies (Canadian and HKD).
• Image Analysis (Where's Waldo): This test resulted in a hallucination, with the model incorrectly stating Waldo was not present in the image. This is presented as a cautionary example.
• Video-Based App Creation: The most impressive demonstration involved uploading an explainer video of an earthquake simulation and prompting the model to create an interactive earthquake visualization app for Japan. The resulting app featured a map of Japan, adjustable earthquake parameters (magnitude), and a visual ripple effect simulating earthquake impact on cities. Crucially, the model inferred all app details from the video itself, without explicit textual instructions.
• Personalized Educational Content: The model generated a chemistry course for kids with lessons, images, and interactive exercises. Some issues were noted with image loading in one lesson.
• Game Development: A prompt to create a 2D platformer game similar to Super Mario resulted in a functional game with enemies, coin collection, and sound effects, all within a single HTML file.

VII. Model Specifications & Benchmarks

Gemini 3.1 Pro accepts text, images, audio, and video input. Its key feature is a 1 million token context window (approximately 700,000 words, a medium-sized codebase, or over an hour of video). The architecture is based on Gemini 3, representing a marginal improvement over previous versions.

Benchmark Results:

• Humanity’s Last Exam: Gemini 3.1 Pro achieved the highest score without tool use, demonstrating extensive world knowledge.
• ARC AGI 2: Gemini 3.1 Pro significantly outperformed other models, including Opus 4.6, showcasing an emergent ability to learn and apply new patterns. The benchmark tests visual puzzle solving and pattern recognition.
• GPQA Diamond: Gemini 3.1 Pro dominated this benchmark, testing graduate-level scientific knowledge.
• Terminal Bench & Agent Coding Benchmarks: Gemini 3.1 Pro also performed well in these areas.
• Long Context Performance: The model maintains accuracy even with extremely long prompts (up to 700,000 words).

Independent evaluations confirm Gemini 3.1 Pro’s intelligence, ranking it as the most intelligent model currently available.

VIII. Cost Efficiency & Hallucination Rate

Gemini 3.1 Pro is not only performant but also cost-efficient, being cheaper than Claude Opus, GPT-5.2, and Gro 4. The model exhibits a lower hallucination rate compared to these competitors, although hallucinations still occur. GLM5, an open-source model, currently has the lowest hallucination rate. The presenter clarifies that a 50% hallucination rate on a specific benchmark doesn’t mean the model is wrong 50% of the time, but rather that it failed 50% of the questions on that particular test.

IX. Conclusion & Resources

The video concludes that Gemini 3.1 Pro is currently one of the most intelligent and performant AI models available. The presenter encourages viewers to try the model and share their experiences. A free weekly newsletter is promoted for staying up-to-date on AI developments, and a link to a HubSpot guide ("The Marketer's Guide to Google Gemini and Notebook LM") is provided for leveraging AI for research and information gathering.