Google didn’t just show up to I/O 2026 — it showed up with receipts. After a year of playing catch-up narratives in the press (fairly or not), the company used its annual developer conference to put Gemini front and center across nearly every product it makes. The result was one of the densest keynotes in recent memory, and if you missed it or blinked during the two-hour runtime, there’s a real chance you skipped past something that will matter to your work within the next six months. Google has since published a highlight reel of the 12 biggest keynote moments, which is a useful anchor — but the highlight reel doesn’t tell you what any of it means. That’s what this piece is for.
How Google Got Here: A Year of Building in Public
Cast your mind back to I/O 2025. Google was in a genuinely awkward spot. Gemini 1.5 had impressed on benchmarks, but the consumer rollout had been choppy, the Gemini app felt unfinished compared to ChatGPT, and the company was visibly scrambling to integrate AI into Search without breaking the ad business that funds everything else. The story the press told — right or wrong — was one of a company reacting rather than leading.
Twelve months later, the picture is different. Google has shipped more Gemini updates in that window than most companies ship in three years. The model family has expanded, the API has matured, and enterprise adoption has accelerated. By the time I/O 2026 rolled around on May 20th, there was actual substance to announce rather than just roadmap promises. That context matters when you’re evaluating whether the keynote represented genuine progress or polished marketing theater. Honestly? It was mostly the former.
The competitive backdrop is also sharper now. OpenAI launched GPT-5 earlier this year, Anthropic’s Claude 4 has been making noise in enterprise coding workflows, and Meta’s open-source Llama push continues to pressure everyone on pricing. Google needed I/O 2026 to land. By most accounts, it did.
Breaking Down the 12 Moments
Rather than walking through the highlight reel in order, it’s more useful to group the announcements by what they actually signal about Google’s strategy.
1. Gemini Omni: The Multimodal Bet Gets Serious
The single biggest announcement of the show was Gemini Omni, Google’s most capable model to date. If you want the full technical breakdown, our earlier piece on Gemini Omni’s multimodal architecture covers it in depth. The short version: Omni processes text, images, audio, video, and code natively — not through bolt-on modules, but as a unified model. That’s a meaningful architectural difference from how most multimodal systems are built today.
What struck me watching the demos wasn’t the capability itself — we’ve seen impressive multimodal demos before — it was the latency. Real-time video understanding with sub-second responses is hard. Google made it look almost boring, which is either a sign of genuine maturity or very carefully rehearsed demos. Probably some of both.
2. Gemini 3.5 Flash: Speed as a Feature
Gemini 3.5 Flash is the model that developers will actually integrate most, and Google knows it. Flash is optimized for speed and cost — it’s not trying to beat Omni on complex reasoning. It’s trying to be the default workhorse for high-volume applications where response time and API pricing matter more than frontier-model performance.
The positioning here is clearly aimed at OpenAI’s GPT-4o mini and Anthropic’s Haiku tier. Developer pricing hasn’t been fully published at time of writing, but early signals from the Google AI Studio interface suggest Flash will be aggressively priced — possibly below $0.10 per million input tokens, which would undercut the current market leaders.
3. Agentic Capabilities Across the Board
Google made clear that agentic AI — systems that can take multi-step actions autonomously — isn’t a side project anymore. It’s the organizing principle for where Gemini is going. Several keynote moments focused on agents that can browse the web, write and execute code, manage files, and interact with third-party services without hand-holding. For a deeper look at what this means for the developer community, the full developer breakdown from I/O 2026 is worth reading alongside this piece.
4. Project Astra Updates
Project Astra, Google’s always-on AI assistant research project, got a significant update. The demo showed a user pointing a phone camera at objects and having Astra provide real-time contextual information — remembering what it had seen earlier in the session, reasoning across time. It’s still positioned as a research preview, but the jump in capability from last year was noticeable.
5. NotebookLM Gets a Major Expansion
NotebookLM quietly became one of Google’s most-used AI products over the past year, and I/O 2026 brought a serious expansion. New features include collaborative notebooks, better source citation, and deeper integration with Google Drive. For researchers, students, and knowledge workers, this is one of the most practically useful things Google announced.
6. AI in Search: The Ad Question Gets Messier
Google showed more AI-generated search summaries, expanded answer formats, and tighter integration between Search and Gemini. It all looked impressive. But here’s the thing: the question of how this affects Google’s ad revenue still doesn’t have a clean answer. Gemini-powered ads are part of the puzzle, but the economics of AI-first search are genuinely unresolved. Google is betting it can thread that needle. The bet might pay off. It also might not.
7–12: The Remaining Moments
The rest of the highlighted moments covered a range of products and integrations:
- Gemini in Workspace: Deeper integration into Docs, Sheets, Gmail, and Meet — including real-time meeting summaries and smart reply improvements that actually felt useful rather than gimmicky.
- Android AI features: On-device Gemini processing for certain tasks, improved circle-to-search, and new accessibility features that expand on work like the Chromebook face control tools Google has been developing.
- Imagen 4: Google’s image generation model got a significant quality bump. The photorealism improvements are real — text rendering in particular has improved dramatically, which has been a persistent weakness in diffusion models.
- Veo 3: Video generation with audio. This one is early, but the demo clips were striking. Generating coherent video with synchronized ambient sound is technically hard, and the results shown were better than I expected.
- Jules, the AI coding agent: Google’s answer to GitHub Copilot Workspace and OpenAI Codex. Jules can handle multi-file refactoring tasks asynchronously — you assign it work, it runs in the background, and comes back with a pull request. Whether it holds up in production workflows is a separate question from whether the demo was impressive (it was).
- Google AI Ultra subscription tier: New premium pricing tier that bundles access to the most capable models, higher usage limits, and priority access to new features. Pricing sits at the high end compared to ChatGPT Plus, but the bundle logic is clear for power users already in the Google ecosystem.
What This Actually Means
For Developers
The Gemini API is now genuinely competitive with OpenAI’s. Flash gives you a fast, cheap workhorse. Omni gives you frontier multimodal capability. The agent tooling is real and getting better. If you’ve been defaulting to OpenAI out of habit rather than deliberate evaluation, I/O 2026 is a good moment to re-run that comparison. Pricing, latency, and multimodal capability all look different than they did a year ago.
For Businesses
The Workspace integrations are the story here for most enterprises. Real-time meeting summaries, AI-assisted drafting, and smarter search across internal documents are the kinds of features that drive actual adoption rather than pilots that never scale. Google has distribution advantages that pure-play AI companies don’t — most enterprise users are already in Gmail and Drive every day.
For Consumers
The on-device processing push in Android is significant from a privacy standpoint. Running inference locally means less data leaving your phone, which matters to a lot of people even if they don’t track benchmark scores. Astra, when it eventually ships broadly, could genuinely change how people interact with their phones. That’s not hype — the demos showed something qualitatively different from current voice assistants.
Who Faces Pressure
OpenAI’s consumer and developer positions both face more serious competition after this week than before it. Microsoft, as OpenAI’s primary distribution partner, will be watching Google’s Workspace momentum carefully. And smaller AI-native startups in the productivity space — the Notions and Coda’s of the world — now have to compete with capabilities that are baked into tools their users already pay for. That’s a tough position to be in.
The agentic push is also worth watching relative to what other companies are building — the full breakdown of Google’s agentic strategy explains why this particular bet could define the next two years of the AI market more than any individual model release.
Google still has execution risk — it always does. But I/O 2026 was the clearest signal yet that the company has figured out what it wants Gemini to be. Veo 3 and Jules in particular feel like early-stage products that could look very different — and very capable — by I/O 2027. The real test won’t be the demos. It’ll be whether any of this is running reliably in enterprise environments by Q4.
Frequently Asked Questions
What is Gemini Omni and how is it different from previous Gemini models?
Gemini Omni is Google’s most capable model to date, designed to handle text, images, audio, video, and code as a unified system rather than through separate specialized modules. It’s positioned as a frontier model for complex, multimodal tasks where raw capability matters more than cost or speed.
When is Gemini 3.5 Flash available to developers?
Gemini 3.5 Flash was announced at I/O 2026 and is available through Google AI Studio and the Gemini API. Pricing details are still being finalized, but early access has opened for developers who want to test it in their applications.
How does Jules compare to GitHub Copilot Workspace and OpenAI Codex?
Jules handles asynchronous, multi-file coding tasks and returns pull requests — similar in concept to Copilot Workspace and Codex’s agent mode. Real-world comparisons will take time, but the feature set at announcement is competitive. For context on how enterprise coding agents are being evaluated right now, the Gartner 2026 Enterprise AI Coding Agents Quadrant is a useful reference point.
Is the Google AI Ultra subscription worth it?
That depends heavily on how much you use Google’s productivity tools and whether you need access to frontier models like Omni. For power users already in the Google Workspace ecosystem who want the best Gemini capabilities, the bundle logic is reasonable. Casual users will be fine on free or standard tiers.
