Google I/O 2026 starts tomorrow. The official schedule has the Google keynote at 10:00 a.m. PT on May 19, followed by the Developer keynote at 1:30 p.m. PT.
From the rumor pile, there are quite a few things that sound exciting. A new Flash model, maybe called Flash 3.2 or Flash 3.5, supposedly very cheap and strong at coding. Veo 4, which some people are already comparing against ByteDance’s video models. Something called Spark, which sounds like Google’s attempt at a more practical everyday assistant or agent. And Google’s own schedule also points to updates around Google AI Studio and Google Antigravity, which honestly surprised me a little because I thought they had basically moved on from Antigravity.
To be fair, Spark might be interesting to me. I still have a lot of files sitting in Google Drive. If it can actually help me sort files, archive things, and clean up my Drive every day, that would have real value.
But I have always felt that Google’s problem is not really model capability. It is product capability.
I do not need to say too much about Codex. Among the big three AI coding products, I still think it has the lowest barrier to entry and the best overall value for a lot of work.
Claude is different. If you are on the $20 base plan, Claude Code with the latest top model may not always be enough. But if you use Sonnet, the web app, or even just the Office integration, I still think that $20 plan has its own value.
Gemini is the one I struggle with. Outside of NotebookLM, I currently have a hard time finding the value of my Gemini subscription.
The core experience is still shaky
Gemini’s product surface is actually huge. If you subscribe, you can use it in the web app, in Antigravity, in Gemini CLI, and across different Google tools. But at least in my own usage, calling Gemini 3.1 Pro from Antigravity or Gemini CLI still disconnects too often.
This has been happening since Antigravity first launched. It still happens now.
If it worked reliably, Gemini would still be useful for front-end work. But with the current stability, it is hard to use as a normal daily tool.
The web version has another problem: sometimes the model feels noticeably dumber.
I once asked it to search for a company’s latest financial report. Somehow, it prioritized the 2024 annual report and answered using that older report. When I uploaded a screenshot of the 2026 report, it replied that it had not searched the external internet because in the “real” web, the 2026 report did not exist yet, and that it had only used my uploaded screenshot as a shifted timeline scenario.
Then I pasted the actual link. Only after that did it apologize.
The weird part is that Gemini argues with me on the wrong things. But when I ask serious questions where I actually want pushback, it becomes overly agreeable. Whatever I say becomes the truth. It does not really hold a position.
The fitness-log example made this obvious
I usually ask AI to record my workouts while I am training, then ask Gemini for feedback.
Recently, the feedback has basically become “whatever you said is fine.” It does not really analyze my past training logs. It mostly repeats the previous plan back to me. If I want to adjust something, it accepts the adjustment. But it no longer gives me much useful resistance or advice.
Last month it was not like this. Last month it would still talk through movements, point out patterns, and give suggestions. This month it often feels like it has stopped thinking about the problem.
That kind of opaque quality drop is harder to manage than a clear limitation. If a model is weak in a known way, you can route around it. If it randomly feels dumber, you start losing trust in the whole workflow.
So what do I actually want from I/O?
Google’s official AI session says it will cover multimodal capabilities, media generation, robotics, agents, vibe-coding tools, and open models. There are also sessions around Chrome DevTools for coding agents, Firebase as an agent-native platform, and Antigravity moving from prototype work toward production workflows.
On paper, this is exactly the kind of stack Google should be able to win with.
It has Android. It has Chrome. It has Search. It has Gemini, DeepMind, TPU, Cloud, Firebase, Drive, Docs, and a massive developer ecosystem. If any company should be able to connect AI models into real daily workflows, it should be Google.
But that is also why I find it hard to get excited.
Google is very good at showing a lot of strong pieces. It is less good at making those pieces feel like one stable product that I want to rely on every day.
For this I/O, my only real expectation is simple: if Flash 3.2, Flash 3.5, or whatever the new model is called really is strong, I hope Google can keep its intelligence stable. Keep the output quality stable. Keep the product path stable.
That is what I care about most now.
Not another impressive demo. Not another model chart. Not another pile of entry points.
Just a model and a product that do not suddenly feel noticeably dumber when I need them to do real work.