Google I/O 2025: A Generative AI Playground for Developers

Google’s I/O 2025 conference was a full-on generative AI showcase, with major updates that promise to reshape how developers code, design, and ship software.

And we attended so you didn't have to.

We're rounding up key insights from I/O 2025 today:

• Gemini 2.5 Pro: Google’s flagship LLM, now with deep research mode and massive context windows.

• New generative models: Veo 3 for video, Imagen 4 for images, and Lyria 2 for music.

• AI agents: Jules and Stitch, autonomous helpers for coding and UI design tasks.

• Model Context Protocol (MCP): A new open standard to link your AI tools and developer environment.

• Benchmarks and showdowns: How Google’s models stack up against OpenAI, Anthropic, and Meta.

• Developer impact: What these new tools mean for your workflow and future projects.

Let's dive in.

Gemini 2.5 Pro: Deep Research and long memory#

Google DeepMind’s Gemini 2.5 Pro was the star of I/O’s AI showcase. This latest version of Google’s flagship LLM introduces powerful new capabilities:

• Deep Research mode: It is an experimental enhanced reasoning mode designed for highly complex math and coding problems. When enabled, Deep Research allows Gemini to “think” longer and more rigorously through difficult prompts. Google announced that Gemini 2.5 Pro with Deep Research can tackle advanced math proofs and tricky coding challenges far better than before. This mode echoes OpenAI’s strategy with its o-series models (which perform internal step-by-step reasoning), indicating a focus on chain-of-thought logic for difficult tasks.

• Longer context window: Gemini 2.5 offers an extended memory for context up to 1 million tokens, on par with OpenAI’s latest GPT-4.1 model. In practical terms, 1M tokens is roughly 800,000 words—the model can intake hundreds of pages of text or an entire codebase at once. This long context means Gemini can analyze massive documents or multi-file software projects in a single go, keeping more information “in mind” during a conversation.

Note: Earlier Gemini versions had a 128k token context standard, with up to 1M in preview.

• Performance: Google reports that Gemini 2.5 Pro is now the world-leading model on key leaderboards. It ranks #1 on both the WebDev Arena (a real-time AI coding competition) and the LMArena general LLM benchmark. This suggests Gemini has surpassed others in a variety of tasks. In coding-specific evaluations, a preview of Gemini 2.5 Pro outscored all rivals, including GPT-4, on web development challenges. The table below is taken from the WebDev Arena Leaderboard.

Note:

• The same rank (6) is used for multiple models, which often happens in leaderboards when models have statistically indistinguishable scores within their confidence intervals.

• Arena Score reflects overall head-to-head coding competition strength.

• A 95% CI represents the margin of error, so the “true” average score is likely within that range.

• Security and AI agent readiness: Google also bolstered Gemini with advanced security against prompt injection and added support for the new Model Context Protocol (MCP) to better integrate tools and plugins. Notably, they teased an Agent Mode where you can simply describe a goal and Gemini will execute sub-tasks autonomously. This hints at Google’s vision for Gemini-powered agents handling complex workflows across apps.

In short, Gemini 2.5 Pro emerges from I/O 2025 as a stronger, more “thoughtful,” and more context-aware AI, narrowing the gap with its competitors on both reasoning and coding tasks.

New generative AI models unveiled#

Google didn’t stop at text-based models. I/O 2025 saw the launch of several specialized generative models aimed at different media modalities, each pushing the state of the art in its domain:

Veo 3: AI video with sound and story understanding#

Google introduced Veo 3, a text-to-video model that represents a leap in generative video. For the first time, Google’s AI can generate synchronized audio for those videos, meaning you can prompt Veo 3 with a scene description, and it will produce a video complete with relevant sounds.

• Improved quality: Veo 3 significantly improves visual fidelity over its predecessor (Veo 2) and maintains more coherent motion. It excels at understanding context, as you can feed it a whole storyline or script, and it will try to “bring it to life” as a short film clip. Google reports better handling of real-world physics (objects move naturally) and even lip-sync when generating talking characters.

• Audio generation: When describing a bustling city street, traffic noise and crowd sounds are generated; when describing a conversation, the characters’ speech is produced. This multimodal output (video + audio) is a notable advance, closing the gap between AI-generated video and real, recorded video with sound.

Imagen 4#

Imagen 4 is Google’s latest text-to-image model, and it has set a new bar for image generation quality and versatility. Combining speed with precision, Imagen 4 produces highly detailed images and is more reliably aligned with the prompt.

Key improvements#

• Enhanced image quality and detail: Imagen 4 generates highly detailed, photorealistic images with fine textures and supports resolutions up to 2K across various aspect ratios.

• Better text and typography handling: Imagen 4 excels at generating images with accurate and legible text, making it well-suited for applications like posters, greeting cards, infographics, comic strips, and UI elements where clear typography is crucial.

• Improved prompt adherence: The model demonstrates a better understanding of complex and detailed prompts, leading to more accurate and coherent image generation based on user input.

• Faster generation speeds: While already faster than Imagen 3, Google has announced a future variant of Imagen 4 that is expected to be up to 10 times quicker, enabling rapid prototyping and faster iteration.

• Safety features: Imagen 4 includes built-in safety measures like filtering and data labeling to minimize the generation of harmful content. It also utilizes SynthID technology to embed an invisible digital watermark for identifying AI-generated images.

• Multilingual prompt support: Imagen 4 supports prompts in multiple languages, including English, Chinese (Simplified and Traditional), Hindi, Japanese, Korean, Portuguese, and Spanish, catering to a global user base.

Below, we tested Imagen 4 with the following prompt:

The following illustration was generated:

This illustration is an impressive showcase of Imagen 4’s generative strengths in photorealistic and contextually rich scenes. The market environment feels lively and authentic, with vibrant displays of fruits and vegetables, textured fabrics in the background, and a bustling atmosphere enhanced by vendor and shopper interactions. The attention to detail is particularly notable: the intricate patterns on the fabrics, the natural variety in produce colors, and the realistic reflections and shadows all add to the image’s depth and believability.

A standout feature here is the signage in Spanish. Unlike previous models, Imagen 4 renders the text “FRUTAS,” “OFERTAS,” and “PRECIOS” clearly and legibly across multiple signs, each in the appropriate place and with proper alignment. While some minor typographical inconsistencies may be observed on close inspection (for example, repeated or slightly misshapen letters), overall the model demonstrates a significant improvement in multilingual and prompt-accurate text rendering compared to Imagen 3 or DALL·E 3.

Next, we tried a similar prompt, changing “Spain” to “Pakistan” and the language from “Spanish” to “Urdu.”

The following illustration was generated:

In additional tests using Urdu script, however, Imagen 4 still struggled to produce perfectly natural and accurate Urdu writing. Characters and spacing are more recognizable than before, but not entirely correct. This shows that while Imagen 4 excels at Latin scripts, there remains room for further advancement with complex, non-Latin scripts. However, rendering non-Latin scripts in earlier Imagen versions was impossible, so this is still a step forward!

In summary, this scene highlights Imagen 4’s evolution in generating complex, multicultural, and commercial settings—delivering hyperrealistic textures and lighting and strong handling of Latin script signage that brings the market atmosphere to life. For practical creative projects involving real-world locations or businesses, this level of visual and textual fidelity marks an important step forward.

Lyria 2#

Google’s generative AI push extends to music via Lyria 2, a new model for music composition. It was debuted through the MusicLM Sandbox earlier in the year and is now officially in the spotlight. This model can turn text prompts (or even humming) into rich musical pieces, catering to musicians and creators seeking AI inspiration:

• Powerful composition and styles: Lyria 2 can generate music in a wide range of genres and moods. Google describes it as bringing “powerful composition and endless exploration,” as artists can use it to instantly hear a melody or chord progression based on an idea and iterate.

• Interactive Tools: With I/O 2025, Google expanded access to Music AI Sandbox (a playground for Lyria 2) and launched Lyria RealTime, an interactive generative music model that powers the MusicFX DJ tool in Google’s AI Test Kitchen. Lyria RealTime allows real-time control of music generation.

• Creative Collaboration: Google worked with musicians and producers to refine Lyria 2, ensuring the AI’s outputs can be useful starting points rather than final products. All AI-generated audio from Lyria carries a SynthID watermark (as do images from Imagen and videos from Veo) to identify it as AI-generated to prevent misuse.

Note: Lyria 2 is available via API and Google’s AI Studio so developers can integrate generative music into apps. According to their site, it costs $0.002 per second of audio.

Below is the code in Python given by Google to create music:

Step 1: Set the REPLICATE_API_TOKEN environment variable.

Step 2: Install Replicate’s Python client library.

Step 3: Run google/lyria-2 using Replicate’s API.

These three simple steps will lead to the creation of amazing music.

AI agents: Automating coding and UI tasks#

In addition to standalone models, Google introduced AI agents that operate autonomously on specific tasks. These agents leverage models like Gemini under the hood but are packaged to perform higher-level jobs. Two notable agents are Jules (coding agent) and Stitch (UI design agent).

Jules#

Jules is positioned as an autonomous coder rather than just a coding helper. While tools like GitHub Copilot suggest lines of code as you type, Jules can take on larger coding tasks asynchronously, from start to finish. At I/O 2025, Google announced Jules is moving from limited preview to public beta, available to all developers.

Key characteristics of Jules include:

• Works like a junior developer: Google explicitly notes that “Jules autonomously reads your code [and] performs tasks like writing tests and fixing bugs”. This means that you can assign a task to Jules (for example: “refactor this module for performance” or “write unit tests for all functions in this file”), and Jules will analyze the existing codebase, make the changes or additions, and then propose the edits, all without the developer having to hand-hold the process. This frees developers from boilerplate or repetitive tasks.

• Asynchronous and parallel: Jules runs in a secure cloud sandbox and works asynchronously, meaning that when someone invokes it, they will not have to wait idly; the task will be handled in the background.

• DeepMind tech under the hood: Jules uses Gemini 2.5 (and likely its coding-optimized “Flash” variant) to understand and generate code. Because Gemini has a large context window, Jules can take in an entire code repository and “remember” it while making changes. Jules is available wherever Gemini is available.

Note: Jule is initially being offered through Google Labs and Cloud IDE integrations

By launching Jules, Google is directly targeting the territory of developer productivity tools. Microsoft/GitHub’s Copilot X and OpenAI’s code interpreter have similar aims, but Jules pushes further into autonomy.

Stitch#

In the context of AI agents, Stitch and Jules complement each other:

• Stitch acts as a “UI designer + front-end developer” who takes a high-level description or sketch and produces a working interface.

• Jules acts as a “software engineer” who can perform coding tasks within an existing codebase.

Using them together, a single developer could conceivably have Stitch draft the app’s interface and Jules implement backend logic or API integration, significantly automating the app development pipeline. Google’s messaging around these tools suggests a future where developers focus on high-level design and logic, and delegate the heavy lifting to AI agents.

Stitch in Agent Mode can be thought of as an AI product manager that understands the desired user experience and materializes it. You tell Stitch what you need your app to look and feel like, and it does the rest, even refining its output if you say the word.

It’s currently an experiment (available via the Labs site), but it generated a lot of excitement among developers and designers at I/O.

We tested Stitch with the following prompt:

The following UI was generated:

Note: The lab also gave the code that it used to generate the above UI, which has been hidden for now.

Google also mentioned a broader “Agent Mode” coming to the Gemini Chat app, where Gemini can execute actions (like booking a calendar event or performing web searches) in the middle of a conversation. This aligns with the agentic trend but is more about personal assistant tasks. The coding and UI agents (Jules and Stitch) are more targeted, immediate productivity boosters for professionals.

Model Context Protocol (MCP): The new standard for AI tool integration#

One of the most quietly impactful announcements at Google I/O 2025 for developers was the introduction of the Model Context Protocol (MCP). Even though it didn’t get a lot of stage time, MCP could change how AI models, plugins, and developer tools talk to each other, opening up a whole new class of intelligent, context-aware workflows.

What is MCP?#

MCP (Model Context Protocol) is a new open standard created by Google (with collaboration from partners) for passing rich, structured context between AI models and external tools or plugins. Think of it as an “API language” that lets different systems share live information in a secure, standardized way.

Why is MCP a big deal?#

Before MCP:

• AI assistants (even advanced LLMs) were often unaware of the true working environment, limited to a static prompt or a one-way function call.

• Tool integrations (like code assistants, chat plugins, or search extensions) required custom wrappers for each product, making integration brittle and slow to evolve.

With MCP:

• Context-rich workflows: An agent like Jules can now understand not just the code, but also the state of the development environment, test results, recent file edits, or issue tracker status, and is updated live.

• Multi-tool orchestration: AI can seamlessly coordinate tasks across plugins (e.g., fetch documentation, lint code, trigger a CI pipeline) without manual glue code.

• Secure and standardized: MCP allows developers to safely grant the AI access only to specific data or actions, with clear boundaries and audit trails, which is crucial for enterprise use.

What does MCP enable in practice?#

• Autonomous agents: MCP powers the “Agent Mode” in Gemini and tools like Jules. Instead of a developer prompting, “Please run my tests,” the agent sees failing tests and proactively offers help or fixes.

• Plugin ecosystems: MCP is designed to work like a “plugin bus” that all speak the same protocol. This means less vendor lock-in and easier extensibility.

• End-to-end automation: Imagine a workflow where an AI can take a bug report from a tracker, check recent code changes, suggest a fix, run the tests, and open a pull request; MCP makes these handoffs robust and standardized.

• Cross-product AI: With MCP, developers could create workflows that connect AI agents from different vendors (say, a Google Gemini agent and an OpenAI-powered plugin) within a single environment, fostering true interoperability.

Where is MCP being used now?#

• Jules (Google’s coding agent): It uses MCP to securely interact with developers’ cloud IDE, repo, and project tools.

• Workspace and Vertex AI integrations: MCP is being rolled out across Google’s ecosystem, and Google is pushing for broader industry adoption.

Benchmark results#

To put Google’s claims of leadership in context, let’s look at how Gemini 2.5 Pro stacks up against other cutting-edge models announced or updated recently. We compare across three domains: reasoning ability, coding performance, and multimodal tasks. 

Each table below summarizes benchmark results for Google’s Gemini 2.5 Pro, OpenAI’s GPT-4.1 (and its reasoning-specialized o3 model), Anthropic’s Claude 3 (latest Opus version), and Meta’s newest LLaMA 4 model. All these models debuted in late 2024 or 2025, making this a fairly up-to-date comparison of the state of the art.

Reasoning and knowledge benchmarks#

We use two indicative evaluations here: 

• MMLU (Massive Multitask Language Understanding), a broad test of knowledge and reasoning across 57 subjects. 

• AIME 2024, a challenging math competition (American Invitational Math Exam) used to gauge complex problem-solving. 

MMLU is a good general-purpose metric (higher = better), while AIME specifically stresses multi-step mathematical reasoning.

MMLU scores reflect accuracy on a broad set of questions – GPT-4.1 currently leads here with around 90%, while Gemini 2.5 Pro and Claude 3 are not far behind in the low-to-mid 80s.

On the AIME math reasoning test, OpenAI’s specialized o3 model demonstrates its edge, scoring 87% (almost solving the entire exam) versus GPT-4.1’s 48%. This highlights that O3’s “deep thinking” approach pays off on complex problems. Google hasn’t released an official AIME result for Gemini, but the new “Deep Research” mode is intended for exactly these kinds of tasks. We may see Gemini close the gap in future evaluations. Meta’s LLaMA 4, with its giant Mixture-of-Experts design, reportedly reaches high-80s on knowledge benchmarks as well, approaching the closed-source models’ performance. Overall, GPT-4.1 remains slightly ahead in general knowledge reasoning, but the field is tight at the top.

Coding and software development benchmarks#

For coding capability, we compare model performance on code generation and developer-focused challenges. One common metric is HumanEval (pass@1), which measures if a model can write correct solutions to programming problems on the first try. We also include the WebDev Arena score/rank as an indicator of performance in a live coding competition environment, where Gemini 2.5 Pro has been highly successful.

Google, OpenAI, and Anthropic are neck-and-neck in top-tier coding performance, with Gemini 2.5 Pro and Claude 3 showing slight advantages in head-to-head coding arenas, and GPT-4.1 remaining an all-around coding workhorse (especially with its extensive support in tools like Copilot).

Multimodal and creative task performance#

Finally, we compare the models on multimodal tasks. All the latest models have at least some multimodal capabilities: GPT-4.1 can accept images; Claude 3 can handle images (in limited ways); LLaMA 4 was trained on text and images; and Gemini 2.5, while not openly available in a chat with image upload, underpins Google’s image (Imagen) and video (Veo) generators. Google’s approach has been to use specialized models (Imagen 4, etc.) for generative outputs, but Gemini itself is multimodal in its understanding (e.g., powering features like interpreting charts or screenshots). A useful benchmark here is MMMU (Massive Multitask Multimodal Understanding),

OpenAI's GPT-4.1 leads in image+text reasoning with a 74.8% score on MMMU, outperforming Claude 3 (59.4%). While Google and Meta haven’t released official MMMU scores for Gemini 2.5 and LLaMA 4, they are expected to be competitive, with LLaMA 4 likely around 70%. Google focuses on specialized models like Imagen 4 and Veo 3 for generation alongside Gemini’s understanding capabilities. Meta’s LLaMA 4 integrates multimodality, directly outputting image features and showing potential in early tests for captioning and analysis.

Cost: How do leading models compare?#

Pricing and accessibility are crucial for developers and teams deciding which generative AI model to adopt. While performance and features are important, cost per use can have a significant impact on real-world adoption, especially for startups, indie devs, and researchers.

Below is a comparison of published API and commercial pricing for the top models. Prices are for API usage, and may differ for special enterprise deals or platform-specific packages.

Note: Pricing as of May 2025

*OpenAI has expanded context in “preview” access up to 1M for enterprise customers.

**Token context for LLaMA 4 may vary depending on implementation and hosting provider.

Gemini 2.5 Pro is competitively priced, especially considering its ultra-long context window, which makes it particularly attractive for use cases involving large documents or entire codebases. In contrast, OpenAI’s GPT-4.1 remains the most expensive per token but is also the most widely integrated and often delivers the highest performance for general users. Claude 3 Opus occupies a middle ground: it is less costly than GPT-4.1 and excels in safety and context handling, though it offers a smaller context window compared to Gemini. Meanwhile, LLaMA 4 stands out by being open source and free to use for inference, but users are responsible for their own compute costs, for example, cloud or server fees, if they do not host the model locally.

What did I/O 2025 signal for the future? #

Google I/O 2025’s generative AI announcements signaled practical, career-changing advantages for developers, engineers, and software teams:

• Jules, the autonomous coding agent, can take over repetitive or boilerplate work: writing unit tests, refactoring code, updating dependencies, or generating documentation. Instead of just code completion, Jules analyzes entire codebases, operates asynchronously in the cloud, and can open pull requests, thus freeing the software developers to focus on design, architecture, and problem-solving.

• Gemini 2.5 Pro’s Deep Research mode means software engineers can delegate even the hardest problems (complex logic, algorithmic puzzles, advanced debugging) to AI.

• Stitch bridges design and development, and turns hand-drawn wireframes or text descriptions into production-ready UI code (HTML, CSS, React). No more tedious translation of mockups to frontends: iterate UI/UX directly in conversation with AI.

• Integrated generative tools (Imagen 4, Veo 3, Lyria 2) are now available in the Google developer ecosystem. They can be generated instantly with a prompt.

• With a 1 million-token context window, Gemini 2.5 Pro can “see” and reason over enormous codebases, entire multi-file projects, or huge technical documents all at once. This is a game changer for code search, refactoring, or onboarding new devs.

• The rise of multimodal models (text, image, audio, video) means software engineers can build applications that see, hear, and generate creative content. For example, they can:

  • Create AI-powered video or image editors.

  • Build tools that automatically generate UI assets, documentation images, or marketing materials.

  • Integrate smart content generators in their apps, increasing value for users.

• The new agentic models (Jules, Stitch, and the upcoming Agent Mode in Gemini) can take on complex tasks end-to-end. As these agents mature, software engineers can expect to spend less time on manual setup and more time on creative engineering.

• MCP is to AI what HTTP was to the web. It promises to:

  • Break down barriers between tools, letting AI work across the whole developer stack.

  • Accelerate building of agentic, context-aware, and fully automated dev workflows.

  • Make it easier for the developers to create and use AI-powered plugins, regardless of which vendor builds them.

The future is closer than you think#

If there’s one thing that’s clear from this conference, it’s this: the future of software development is going to be quite the collaboration.

Jules and Stitch aren’t here to replace you.

They’re here to take the boring work off your plate. If Google gets it right, the next dev cycle might just be a lot more fun.

So go ahead and start delegating boilerplate and code to AI, so you can focus on the bigger ideas.

As we move forward, you can build your Generative AI skills with Educative's courses: