Honestly, May 2025 felt like the month the "AI bubble" finally stopped being about chatbots and started being about things that actually do work. You've probably seen the headlines, but the vibe on the ground at Google I/O and Microsoft Build was different this time. It wasn't just "look at this poem I wrote." It was "here is an agent that just booked your flight and fought with customer service so you didn't have to."
The month was a whirlwind. We saw Google basically reinvent its search engine—again—while Anthropic and Microsoft dropped updates that make the 2024 versions of these tools look like calculators. If you’ve been waiting for AI to become "useful" rather than just "cool," May 2025 was your month.
Google I/O 2025: The "AI Mode" Takeover
Google didn't just sprinkle some AI on top of Search this year. They basically built a new house around it. The big star was AI Mode. Unlike those "AI Overviews" we’ve been seeing for a while, AI Mode uses something Google calls a "query fan-out technique."
Basically, if you search for something complex—like "plan a 3-day trip to Tokyo that avoids crowds but includes a Michelin-star ramen spot and a place to buy vintage vinyl"—Google doesn't just give you a list of links. It breaks that massive prompt into ten smaller searches, runs them all at once, and stitches the answer back together.
What most people got wrong about Gemini 2.5
People were expecting Gemini 3.0, but we got Gemini 2.5 Pro instead. Don't let the decimal point fool you. This version is built for "Deep Thinking." Google actually launched a specific Deep Think mode that’s supposedly better at high-level math and coding than almost anything else on the market.
They also introduced a new tiering system:
- Google AI Pro: The old $19.99/month plan, rebranded.
- Google AI Ultra: A whopping $249.99/month tier for people who want "early and maximum access" to the bleeding edge.
It’s expensive. Kinda ridiculous, right? But for developers using the new Project Mariner—an agent that can literally browse the web and navigate apps on your behalf—it’s becoming a "cost of doing business" tool.
Anthropic’s Claude 4 Surprise
While Google was busy talking about Search, Anthropic quietly dropped Claude 4 on May 22. It includes Claude Opus 4 and Sonnet 4. If you use AI for coding, this was the real news of the month.
The most interesting feature is "Tacit Knowledge." Anthropic claims these models have improved memory that allows them to extract and save key facts over time. They aren't just forgetting everything as soon as you close the tab anymore. They are building a context of how you work.
Claude Code also went generally available. It now runs background tasks via GitHub Actions. You can basically tell your IDE to "fix the CSS bugs in the footer while I'm at lunch," and it actually happens.
The Microsoft Build "Agentic Web"
Microsoft Build 2025 was all about agents. They are moving away from "Copilot as a sidebar" to "Copilot as a coworker."
The biggest technical shift was the Agent2Agent (A2A) protocol. This allows an AI agent built by one company to talk to an agent built by another. Imagine your personal shopping agent talking directly to a store’s inventory agent to negotiate a price. No human middleman.
Microsoft also showed off Computer Use for Copilot Studio. It’s exactly what it sounds like. The AI can now "see" your desktop and move the mouse. It’s a bit creepy, honestly. But it means it can automate things in old software that doesn't have an API, like that weird 20-year-old accounting program your office refuses to replace.
Apple Intelligence and the Privacy Pivot
Apple didn't wait for June's WWDC to start leaking the good stuff. In May, we saw the first real look at Visual Intelligence for screenshots.
In the new iOS, you can take a screenshot of literally anything—a concert poster, a weird error message, a pair of shoes—and the AI analyzes it instantly. It can even pre-populate your calendar from a screenshot of a flyer.
They also announced Workout Buddy for the Apple Watch. It’s an AI voice that uses your real-time heart rate and history to give you actual coaching during a run. It’s not just a "keep going" recording; it’s analyzing your gait and telling you to slow down if your form is slipping.
📖 Related: Where is AirDrop on the iPhone? How to Find and Use Apple's Best Sharing Tool
What’s Next: Actionable Steps for the Post-May World
The "wait and see" period for AI is officially over. Here is how you should actually be using these updates:
- Audit your "Deep Research" needs: If you're still doing 50 manual Google searches for one project, try the new AI Mode (rolling out to all US users as of May 20). It’s specifically designed to handle "research debt."
- Look into MCP (Model Context Protocol): If you're a developer or a tech-heavy business, this is the new standard for how AI talks to your data. Microsoft and Anthropic are both all-in on this.
- Test the "Agent" Workflow: Stop asking your AI to "write an email." Start asking it to "check my last three emails from Dave, find the spreadsheet he mentioned, and summarize the budget rows." The models from May 2025 can actually handle that level of multi-step reasoning.
We are seeing a shift from AI as a "content creator" to AI as an "operator." The tools are finally getting hands. Whether we're ready for them to start clicking buttons for us is another question entirely.