The Gemini Stack Just Got a Major Upgrade
From a brand-new Gemini 3.5 Flash model to Antigravity 2.0, the Antigravity CLI, the Omni world model, and the Gemini Spark agent — here is everything Google announced and why it matters.
If you blinked during Google I/O 2026, you missed roughly half a roadmap. On May 19–20, 2026 Google delivered one of its densest developer keynotes in years — repositioning Gemini from a chat-first product into a full agent operating layer that spans the CLI, the desktop, mobile, Workspace, Android and Google Cloud. This post unpacks every major announcement, with infographics, comparisons, and a practical migration view for developers.
1. Gemini 3.5 Flash — Frontier Intelligence, at Flash Speed
The headline model is Gemini 3.5 Flash. Google's pitch: frontier-level quality at a fraction of the cost, with output speeds up to 4× faster than competing frontier models. According to Google, 3.5 Flash outperforms Gemini 3.1 Pro across nearly all benchmarks — coding, agentic tasks, and multimodal reasoning — while being cheaper and dramatically faster. CEO Sundar Pichai called it "remarkably fast" in his press briefing.
Where 3.5 Flash is rolling out today: the Gemini app, Google Search AI Mode, Antigravity 2.0, and the Gemini API. Google also said it strengthened cybersecurity protections to reduce harmful outputs and reduce false refusals on safe prompts — a quietly important tuning step. Gemini 3.5 Pro, the heavier sibling, is still being tested internally and is expected to launch publicly the following month.
The strategic move here is interesting: by anchoring the default tier on Flash rather than Pro, Google is optimizing for the use case it sees as the next bottleneck — concurrent agent fan-out. When five subagents are reasoning in parallel, model latency compounds. A 4× speed advantage per call becomes a 4× advantage per pipeline.
2. Antigravity 2.0 — Goodbye IDE, Hello Agent Orchestrator
The most architectural change is Google Antigravity 2.0: a brand-new standalone desktop app that is no longer an IDE-style assistant. It's an agent-first workbench built around parallel agents, scheduled tasks, dynamic subagents, and persistent background workflows.
The unifying idea is the agent harness: a single core that powers all five surfaces. Whatever Google improves in the harness lands in every surface automatically. Antigravity 2.0 also keeps the four building blocks developers loved from Gemini CLI — Agent Skills, Hooks, Subagents, and Extensions (now rebranded as Antigravity Plugins).
Managed Agents in the Gemini API
This is the sleeper announcement. With a single API call, you can spin up a fully isolated Linux environment that can reason, call tools, and execute code — and the environment persists across turns, with files and state intact. Three pieces:
- Agent harness: the same internal infrastructure that powers Google's own agents, co-optimized with Gemini 3.5 Flash.
- Persistent isolated environments: each interaction creates a sandbox you can resume — no reinitialization of context.
- Custom agent definitions: extend the default agent with your own skills and markdown instructions; templates are available in the Google AI Studio Playground.
3. Antigravity CLI vs. Gemini CLI — The Migration You Need to Know
Here's the part that's already causing friction with the developer community. According to Google's Cloud release notes and a GitHub announcement, the Gemini CLI and Gemini Code Assist IDE extensions will stop serving requests for Google AI Pro, Google AI Ultra, and the free individual tier on June 18, 2026. Affected users are being directed to migrate to the new Antigravity CLI.
Enterprise customers using Gemini Code Assist Standard or Enterprise, Google Cloud access, or paid Gemini / Gemini Enterprise Agent Platform API keys will retain Gemini CLI access. Google has said it will continue maintaining the Gemini CLI Apache 2.0 repository with model updates, bug fixes, and security patches for those enterprise users. The Register has noted that Antigravity CLI is not being released as an open-source repo in the same way Gemini CLI was — a change that has generated visible community pushback.
| Capability | Gemini CLI (legacy) | Antigravity CLI (new) |
|---|---|---|
| Default model | Gemini 3.x family | Gemini 3.5 Flash |
| Implementation | Apache 2.0, open source | Go binary, proprietary distribution |
| Execution model | Mostly synchronous | Async, multi-agent, background-friendly |
| Agent Skills, Hooks, Subagents | Yes | Yes |
| Extensions / Plugins | Extensions | Antigravity Plugins (rebrand) |
| Shared harness with desktop | No | Yes (same core as Antigravity 2.0) |
| Free tier access | Cut off June 18, 2026 for AI Pro / Ultra / free users | Free + paid tiers (AI Ultra at $100/mo offers 5× higher limits) |
| Enterprise availability | Continues for Gemini Code Assist Std/Ent, GCP, paid API | Enterprise Agent Platform integration |
| Feature parity at launch | N/A | Google warns: "not 1:1 right out of the gate" |
4. Omni — Google's New World Model for Editable Video
Gemini Omni is Google's new family of multimodal models built around reasoning and generation. Omni accepts text, image, audio, and video as inputs, and produces editable, real-world-knowledge-grounded video. In Google's own words: "You can combine images, audio, video and text as input and generate high-quality videos grounded in Gemini's real-world knowledge. You can also easily edit your videos through conversation."
The differentiator vs. peer products (Sora, Runway, Adobe Firefly Video, etc.) is the world-model lineage. DeepMind has spent years on world models for robotics — systems that predict how environments change based on actions. Omni inherits that DNA, which means it leans on physical plausibility, not just prompt-to-pixel generation. Examples Google highlighted: take an existing video and ask Omni to change what's happening, add a new character, or modify an object.
5. Gemini Spark — The General-Purpose Agent
Gemini Spark is Google's new general-purpose AI agent that can reason across connected apps and act on the user's behalf. Spark launches in beta the week after I/O for trusted testers and Google AI Ultra subscribers. The vision: turn Gemini from a chatbot into an active assistant that operates across Search, Android, YouTube, and Workspace.
This is the consumer-facing twin of the Antigravity story. Where Antigravity is for developers building agents, Spark is the assistant users talk to while it quietly does work behind the scenes.
6. AI Studio Goes Mobile, Workspace-Native, and Android-First
Google also reshaped Google AI Studio into a more end-to-end product:
- AI Studio mobile app — pre-register this week; capture ideas on the go and pick them up on desktop.
- Export to Antigravity — one click moves an entire AI Studio project, with full context, into local Antigravity.
- Workspace integration — agents can natively call Google Workspace APIs (Docs, Sheets, Calendar) and embed them in apps.
- Android support — build Android apps from a prompt.
- Google Play Console integration — publish to the test track without leaving AI Studio.
Putting It All Together — One Picture
What This Means for Developers
Read between the announcements and three themes emerge:
- Agents are the new primitive. The IDE assistant era is fading. Google is now optimizing for workflows that run themselves — scheduled, parallel, persistent.
- Flash is the new default. Pricing and latency are now first-class design constraints. Google is betting most agentic workloads do not need a top-shelf reasoning model on every call.
- Openness is being rebalanced. Gemini CLI was open source; Antigravity CLI is not, at least not yet. Expect the community-fork discussion to continue.
If you build with Gemini, the immediate action items are concrete: try Antigravity CLI, audit your free-tier Gemini CLI usage before June 18, 2026, experiment with Managed Agents for any workload that needs an isolated sandbox, and start thinking about which of your existing Gemini calls should switch to 3.5 Flash for the latency boost.
Google I/O 2026 keynote. Gemini 3.5 Flash, Antigravity 2.0, Antigravity CLI, Omni, and Spark announced.
Gemini Spark beta opens for trusted testers and Google AI Ultra subscribers.
Gemini 3.5 Pro expected to launch publicly after internal testing.
Gemini CLI & Gemini Code Assist IDE extensions stop serving Google AI Pro, Ultra, and free individual tiers. Migrate to Antigravity CLI.
Closing Thoughts
Google I/O 2026 was less about a single hero model and more about a refactor of the entire Gemini surface area. The center of gravity moved from "chat with a smart model" to "orchestrate a fleet of agents on a shared harness." Whether you cheer the speed and capability gains or grumble about the loss of an open-source CLI, one thing is clear: if you build with AI, your toolchain is about to change.
Sources: Google I/O 2026 keynote and blog posts; Google Cloud release notes; GitHub announcement by Dmitry Lyalin; MarkTechPost (May 19, 2026); Tech Startups (May 20, 2026); Let's Data Science (May 20, 2026); The Register coverage. All product names are trademarks of their respective owners.