Google I/O 2025: AI Mode, Veo 3, and the search engine that thinks
Google I/O 2025 had a different energy from the previous year. In 2024, the company had stumbled through its Gemini launch with several high-profile errors and a demo video that turned out to be sped up. This year the announcements were more confident, the demos were more live, and the products were more clearly close to shipping. The headline items were AI Mode in Google Search, Veo 3 video generation, and Project Astra updates. Alongside all of it, Anthropic announced that Claude 4 had completed a seven-hour autonomous coding session without human intervention.
AI Mode: search gets a complete redesign
AI Mode is the most significant change to Google Search since the introduction of Featured Snippets in 2014. Instead of returning a list of links with a short summary, AI Mode interprets your query, performs multiple sub-queries in the background, synthesises the results, and presents a conversational response with citations. It is closer to a research assistant than a search engine.
The implications for the web are substantial. Google has always been the primary traffic source for publishers and content sites. AI Mode reduces the need to click through to source material for informational queries. Publishers have been watching Search AI Overviews, the smaller predecessor feature, nervously for a year. AI Mode is a much more significant step in the same direction.
Google's argument is that AI Mode drives more clicks for complex queries because it helps users understand what they are actually looking for before they commit to a source. The early data is contested, and the advertising model for AI-generated responses is still being figured out.
Veo 3: video with synchronised audio
Veo 3 is Google's video generation model, and it introduced a capability that previous systems did not have: synchronised audio generation. You can prompt the model to generate a short video clip with dialogue, ambient sound, and sound effects all rendered in sync with the visuals. Earlier video models like Sora produced silent video, leaving audio as a separate post-production step.
The demos at I/O were genuinely impressive. A prompt describing a market scene in a specific city produced a video with background conversation, footsteps, and ambient market noise. The synchronisation between mouth movements in human figures and generated speech was noticeably better than what had been publicly available before.
Project Astra: persistent AI assistants
Project Astra is Google's research prototype for a persistent AI assistant that can see, hear, and maintain context over time. The I/O 2025 demos showed the assistant running on a phone, navigating a room, and remembering earlier parts of a conversation to inform later responses. It could identify objects, read text in the environment, and answer questions grounded in what it was seeing in real time.
The "persistent memory" aspect is what distinguishes Astra from a standard voice assistant. Rather than treating each query in isolation, it maintains a context window that spans the current session and can reference earlier observations. The practical applications in accessibility, field work, and personal productivity are fairly obvious.
Claude 4 works autonomously for seven hours
Anthropic released Claude 4 Sonnet and Claude 4 Opus in May 2025. The announcement that attracted most attention was a benchmark result: Claude 4 Opus had completed a software engineering task autonomously for seven consecutive hours without requiring human feedback or course correction.
This was not a benchmark on a standardised dataset. It was a report from an enterprise customer who had deployed Claude in an agentic workflow. The model maintained context, tracked its own progress, identified when it had made mistakes, and corrected them without external prompting. Seven hours of uninterrupted autonomous work represents a meaningful threshold for practical deployment in software development pipelines.
The broader pattern
Google I/O 2025 illustrated something about where the AI product cycle is heading. The first wave of AI products, from 2022 to 2024, was about demonstrating capability. Chatbots that could write code, summarise documents, pass exams. The second wave, now arriving, is about integrating those capabilities into the infrastructure people already use, in ways that are persistent, multimodal, and increasingly autonomous.
AI Mode changes the experience of using search. Veo 3 changes what a small team can produce in video. Astra changes what an assistant can do over time. Claude working for seven hours changes how software gets built. None of these are incremental improvements to existing features. They are category changes.