Remember when GPT-4 dropped in early 2023 and we all thought we’d peaked? I was sitting in my home office, halfway through building a customer support bot that could barely handle multi-turn conversations, and suddenly the ceiling lifted. That feeling of “okay, everything I’m building just changed” — I got hit with it again this May. Except this time, it wasn’t one announcement. It was a cascade.
Google’s Agentic Play With Gemini 3.5 and Gemini Omni
Google’s May 2026 updates landed with a clear thesis: we’re in the “agentic” era now. Their new Gemini 3.5 model and the companion Gemini Omni are built for advanced reasoning and creation — two words that matter enormously when you’re designing bots that need to do more than retrieve answers from a knowledge base.
From my workbench, here’s what this means practically. Gemini 3.5’s reasoning capabilities suggest we can build agents that plan multi-step workflows with less hand-holding from developers. If you’ve ever spent days writing elaborate chain-of-thought prompts just to get a bot to handle a three-step booking process, you know the pain. Stronger native reasoning reduces that scaffolding.
Gemini Omni is the piece that excites me most for the multimodal bot architectures we discuss on ai7bot.com. A model designed for creation alongside reasoning opens up agent pipelines where your bot doesn’t just understand a request — it produces artifacts. Think code generation agents, design assistants, or bots that draft and iterate on documents within a conversation flow.
Google had already been leaning into agentic AI at Cloud Next ’26 in April, so May’s announcements feel like the product delivery on promises made a month earlier. For those of us building on Google’s stack, the signal is clear: architect your bots as agents, not chatbots.
OpenAI’s Real-Time Voice and Translation Models
On the OpenAI side, May brought real-time voice and translation models specifically designed for AI agents. This is a different kind of upgrade — it’s not about smarter reasoning, it’s about faster, more natural interaction layers.
I’ve been building voice-enabled bots for about eighteen months now, and the latency problem has always been the killer. You can have the smartest agent in the world, but if there’s a two-second pause between the user speaking and the bot responding, the experience falls apart. Real-time models attack this directly.
The translation angle is equally significant for anyone building bots that serve multilingual users. Previously, you’d chain a transcription model, a translation step, your main LLM call, another translation step, and then text-to-speech. Each hop added latency and potential error. Models purpose-built for real-time translation in an agentic context could collapse that pipeline considerably.
What This Means for Your Bot Architecture
Here’s my take as someone who spends every day in this space, wiring up agents and debugging tool-calling chains:
- Plan for agentic patterns now. If you’re still building simple request-response bots, you’re leaving capability on the table. Both Google and OpenAI are optimizing their newest models for multi-step, tool-using agents. Your architecture should reflect that.
- Voice is no longer a nice-to-have. With real-time models entering the picture, voice-first bots become viable for production use cases that previously demanded human operators. Start prototyping voice agent flows even if your current projects are text-only.
- Multimodal creation changes the output layer. Bots that can generate images, code, documents, or structured data as part of their workflow open up entirely new product categories. Consider what your bot could produce, not just what it can answer.
- Watch the cost curve. More capable models often start expensive. Plan your architecture with fallback routing — use lighter models for simple queries and reserve the heavy reasoning for tasks that demand it.
My Honest Reaction
May 2026 felt like a month where the major labs aligned on a shared vision — agents over chatbots, real-time over batch, creation over retrieval — and shipped products that back it up. For those of us building smart bots daily, this is both thrilling and humbling. The tools are advancing faster than most of us can integrate them.
My plan? Pick one new capability — probably Gemini Omni’s creation features — and build a focused tutorial around it for the ai7bot.com community. That’s how I process these shifts: by building something concrete with them. I’d encourage you to do the same.
🕒 Published: