The Other Side of Google I/O 2026 — What Disappointed and What Surprised
Parts 1 through 4 covered what Google got right: Flash economics, serverless agents, Gemini Omni, and the Gemini CLI shutdown. This final piece covers the other side — three things that disappointed and one thing that surprised everyone, including me.
Disappointment 1: Antigravity 2.0 — Great Vision, Brutal Execution
The demo was spectacular. A demo showing an OS built by 93 agents over 12 hours, plus a playable Doom clone, all on stage. Agents as first-class deployment targets — versioning, rollback, observability baked in. The vision is sound.
Then the forced update shipped.
No migration path. No opt-out. Developers who were mid-sprint woke up to broken builds. Existing projects that worked on Monday stopped compiling on Tuesday.
The Google AI Forum filled up fast. Thread titles include "Incompetence 10.0" and "The WORST IDE for development." Source control and the local terminal — tools most developers consider non-negotiable — were missing from the default installation. Users had to manually hunt through settings to re-enable them.
This is the part that stings. Source control is not a power-user feature. A local terminal is not optional. When these are absent from a default install, it reads like the product was built by people who do not ship code themselves. Or worse — it reads like a team optimized for the demo reel and forgot the people who would use the product on Wednesday.
When your platform update breaks the projects people built on your previous platform update, the 12-hour demo stops mattering.
Disappointment 2: Gemini Spark — A Gmail Bot, Not an Agent
Google pitched Gemini Spark as a 24/7 autonomous agent. What shipped: Gmail, Google Docs, Sheets, Slides, Google Calendar. Plus a handful of consumer integrations — Canva, OpenTable, Instacart.
No Slack. No GitHub. No Linear. No Jira. For any developer whose workflow lives outside Google Workspace, Spark has zero surface area.
Here is the irony. The same company, on the same day, shipped the Managed Agents API with dozens of external integrations — GitHub, Jira, Stripe, Linear, Notion, MongoDB — plus native MCP support and Claude model compatibility. The API side built a genuinely open platform. The consumer product stayed inside Google's walled garden.
Same company. Same keynote. Opposite directions.
If the Managed Agents API team and the Spark team compared notes, you would not guess they work in the same building.
Now compare Spark to what already exists. Claude Code and OpenAI Codex both operate inside your actual development environment — your files, your terminal, your version control. Spark operates inside Google's apps. The conceptual overlap is real, but the practical utility gap is wide. Spark does not meet developers where they already work. It asks them to move into Google's house first.
Disappointment 3: Android 17 — The Invisible OS
Android was barely mentioned during the two-hour keynote. At Google I/O — the event that used to be about Android.
The major Android announcements? They happened the week before, off-site, at a pre-recorded event called "The Android Show." The features that did reach the I/O stage — smarter notifications, on-device summarization — are incremental. No new design language. No major API surface. No surprise.
The entire two-hour keynote was Gemini, AI, agents. Android felt like an extra in someone else's movie.
The read is straightforward: Google is betting its future on AI services, not on the operating system that carries them. Whether that is the right call depends on your vantage point. But doing it at I/O — in front of the Android developer community, the people who build the apps that make Android worth using — felt like Google forgot to impress the people in the room.
The Bright Spot Nobody Expected: Project Aura
In the middle of a keynote dominated by software, Google showed hardware. And it might have been the best thing on stage.
Project Aura is a collaboration between Google and XREAL — Android XR glasses that split the compute off the face. The glasses weigh roughly 80-90g. The heavy lifting happens in a tethered puck running a Snapdragon XR2+ Gen 2 chip. The result: an OLED display with a 70-degree field of view, wide enough to show three apps side by side, in a frame light enough to wear for hours.
Where Aura Sits
| Meta Ray-Ban | Project Aura | Apple Vision Pro | |
|---|---|---|---|
| Weight | 52g | 80-90g (glasses only) | 750g |
| Display | None | OLED, 70-degree FOV | Micro-OLED, 90-degree FOV |
| Apps side by side | N/A | 3 | Unlimited (spatial) |
| Compute | On-frame (Snapdragon AR1) | Tethered puck (XR2+ Gen 2) | On-device (M2 + R1) |
| Text readability | N/A | Sharp (hands-on reports) | Excellent |
| Price | $299 | TBA (expected well below $3,499) | $3,499 |
| Primary use | Camera + audio | AR workspace + media | Spatial computing |
Meta Ray-Ban is light but has no display — it is a camera and speaker on your face. Apple Vision Pro has an extraordinary display but weighs 750g and costs as much as a laptop. Aura lands in the gap: actual visual output, actual wearable weight, at a price that is expected to undercut Vision Pro by a significant margin.
Hands-on reviewers at I/O reported text was sharp and pixels were not visible — a meaningful upgrade from the CES 2026 prototype shown earlier this year. But the demo that drew the most attention was not the specs. It was this: connect the glasses to a laptop via DisplayPort, and they become a virtual large monitor. No physical screen. No desk. Multiple reviewers called it the most practical demo of the entire event.
Google also announced a Developer Catalyst Program giving developers early access to devkits. AR/XR glasses live or die on the app ecosystem. Hardware without software is a paperweight. Getting devkits into hands early is the right move.
The broader signal is what makes Aura interesting beyond the product itself. AI has been a software story — models, APIs, tokens, agents. Aura is AI expanding into the physical interface layer. If a developer can carry a full workspace in a glasses case — no monitor, no desk, no office — the implications for remote work and mobile development go beyond what a new model release can offer.
Global launch is targeted for 2026. No price yet. That is the one caveat worth watching — if the price lands above $1,000, the sweet-spot argument weakens considerably.
The Series in Five Lines
- Part 1: Flash 3.5 vs Pro — Pro Performance, Flash Branding — The "cheap model" costs 15x what Flash cost two generations ago.
- Part 2: Managed Agents API — Serverless Agents Are Here — Deploy, scale, monitor. One CLI command.
- Part 3: Gemini Omni and the Gemini CLI Shutdown — The best demo and the worst goodbye, on the same day.
- Part 4: What the Numbers Actually Say — Pricing deep-dive and the open-source burial.
- Part 5: What Disappointed and What Surprised — You are here.
Four wins. Four misses. One hardware surprise. All announced in the same 48 hours.
This wraps the Google I/O 2026 series. If you sat through the keynote live or tested any of these products hands-on — what surprised you? Drop a comment or find me on GitHub.
Top comments (0)