Sebastian Gomez

Jun 24, 2026

Updated: Jun 25, 2026

The Gemini universe at Google I/O: powering your projects with the most advanced AI

If there was one thing that electrified me at this year's Google I/O, it was the session dedicated to the Gemini model ecosystem. I was there live, listening to Luciano Martins (AI Developer Advocate and TLM, Google DeepMind) and Shrestha Basu Mallick (Group Product Manager for the Gemini Developer API), and I was absolutely blown away by what they shared.

Context note: This post is a recap of the session as it was presented at Google I/O in May 2025. The benchmark numbers and ranking positions are a snapshot of that moment. Gemini models evolve very quickly, so treat them as a historical reference.

Ever since Google launched Gemini in late 2023, it has been clear that this is no ordinary AI model. The key to its brilliance is that it is multimodal from the start. This means that, unlike other models that were adapted afterward, Gemini was built from the ground up to handle and understand an incredible variety of information formats. Not just text, but also images, audio, video, and even code. This opens up a world of possibilities to build truly innovative solutions.

A look at the Gemini ecosystem: a model for every need

Luciano and Shrestha gave us a tour of the different Gemini model families, each one optimized for specific use cases. It is amazing how Google has thought about the scale and the diversity of the needs we developers have.

• Gemini 2.5 Pro: This is the most powerful model in the portfolio. Designed for extremely complex tasks that require deep reasoning and "thinking" capabilities. If your application needs maximum intelligence, this is your option. (At Google I/O 2025 it was presented in preview, today the 2.5 family is already generally available.)
• Gemini 2.5 Flash: Do you need excellent general performance across many tasks? Flash is the answer. It is a fast, efficient model that optimizes the cost and performance balance. Perfect for agility. (Also in preview at the time.)
• Gemini 2.0 Flash Lite: As its name suggests, it is a lightweight model, optimized for speed and efficiency. Ideal for high volume use cases, such as quick summaries or text classification.
• Gemini Nano: AI in your pocket! This model is designed specifically to run directly on the device (on device), such as on Android smartphones. It lets applications have advanced AI capabilities without relying on the cloud, improving privacy and latency.
• Gemini Embedding: If your application needs to organize information at scale or perform semantic search, this is the model. It turns text into high quality vectors that capture its meaning and relationships, making it easier to understand and organize data.

You can review the features and limits of each model in the official Gemini models documentation.

Note: Gemini model names and availability change frequently; some 2025 references may have been renamed or retired. Check the official models documentation for the current catalog.

It is worth noting that, although earlier versions such as Gemini 2.0 Flash and 1.5 were still accessible at the time, Google strongly encouraged us to migrate and experiment with the 2.5 models. Their improved capabilities and performance will surprise you!

Gemini dominating the benchmarks

But how did Gemini compare to the competition at Google I/O 2025? This is where the presentation got really exciting. At that moment, the Gemini models were proving to be leaders across several key benchmarks:

• LMArena (Chatbot Arena), community preference: On LMArena, an anonymous platform where we developers test different AI models and vote for our preferences, three Gemini models were in the TOP 10. And the most impressive part: in the May 2025 ranking, Gemini 2.5 Pro stood at number 1. This was not just a technical win, it was a win in the preference of the developer community. As Shrestha put it: "We are number one!".
• WebDev Arena (coding): For those of us who build apps and write code, this is music to our ears. In May 2025, Gemini 2.5 Pro also led WebDev Arena, a real time coding competition. If you are looking for a model that helps you go "from zero to one" in building applications or in coding tasks, Gemini 2.5 Pro is a great ally.
• Academic benchmarks: Beyond user preference, Gemini 2.5 Pro also stood out on rigorous academic benchmarks. It led in areas such as reasoning and knowledge (for example, on Humanity's Last Exam, HLE), mathematics (AIME 2024 and 2025), code generation and editing, factuality, visual reasoning, and long context understanding. Even in multilingual performance it proved to be a robust and versatile model.

Note: positions on LMArena and WebDev Arena rotate constantly. The figures above describe the state in May 2025, not necessarily the current situation.

My personal reflection

As a developer, my experience in the Google I/O talk was eye opening. Seeing Luciano and Shrestha present these tools with so much enthusiasm and professionalism confirmed for me that we are at a decisive moment for AI. The speed at which Google is innovating and democratizing access to models as powerful as Gemini is simply astonishing.

I feel incredibly motivated and eager to experiment with all these new capabilities to build impactful solutions. The idea of having AI this powerful, efficient, and accessible, that also understands multimodality, is a true game changer. It is an honor to be part of this community that is building the future of AI.

Suggested exercises

1. Create an account in Google AI Studio and try the same prompt in Gemini 2.5 Pro and in Gemini 2.5 Flash. Compare the quality of the responses and the response times.
2. Take advantage of Gemini's multimodal nature: send it an image and ask it to describe it or to extract data from it.
3. Visit LMArena and check the current ranking. Compare how it has changed relative to the May 2025 snapshot we describe here.

3-point summary

1. Gemini is multimodal from the start, it was built from the ground up to understand text, images, audio, video, and code.
2. The ecosystem is organized by need: Pro for deep reasoning, Flash for speed and cost, Flash Lite for high volume, Nano for on device, and Embedding for semantic search.
3. At Google I/O in May 2025, Gemini 2.5 Pro led LMArena and WebDev Arena and stood out on academic benchmarks such as HLE and AIME, remember those positions change over time.

That is all. I hope this recap is useful to you and that it encourages you to experiment with the Gemini models in a project you have in mind.

So, have you tried the Gemini models yet? Tell me about your experience in the comments, I would love to know what you are building. And remember, if you liked it, you can also share it using the social links below.