Google I/O 2026: The Complete AI Stack

The Model Roster

Google's I/O 2026 didn't introduce a single new flagship — it introduced a coordinated family. The structure: one default model that runs everywhere, one world-model media engine, four updated generative siblings, an open-weights line for on-device, and two agent runtimes. Every announcement maps back to one of these six lanes.

Gemini 3.5 Flash — The Centerpiece

Google's framing: a Flash-tier model that "rivals large flagship models at speeds you expect from the Flash series." The benchmarks back the marketing on coding and agentic work, but with sharp limits on long-context and pure reasoning. Read both columns before flipping production code.

Specs and Pricing

Benchmark Wins vs Gemini 3.1 Pro

Where 3.5 Flash Still Loses to 3.1 Pro

Honest read: 3.5 Flash is not a clean replacement for the Pro tier. It trails on three benchmarks that matter for specific workloads:

• Humanity's Last Exam — frontier knowledge questions. Stay on 3.1 Pro for hardest research-grade reasoning.
• ARC-AGI-2 — novel-problem abstract reasoning. Stay on 3.1 Pro for puzzle-style workflows.
• 128K MRCR v2 — long-context retrieval at depth. Stay on 3.1 Pro if you regularly push past 200K tokens of dense context.

The 12x Speed Inside Antigravity — Read Carefully

Via the public Gemini API, Gemini 3.5 Flash runs approximately 4x faster than comparable frontier models. Inside Antigravity 2.0, it runs at approximately 12x speed. Same model. The delta comes from harness co-optimization — Antigravity's agent harness was rebuilt alongside 3.5 Flash, with shared assumptions about prompt structure, tool-call patterns, and context-cache routing.

This is the most strategically important number in the entire I/O launch and the most under-reported one. It is the strongest practical argument for using Antigravity 2.0 as your daily agent runtime rather than building on the raw API. Three times faster, same intelligence, same price.

Price Position Within the Gemini Family

Gemini Omni — The "World Model"

The most architecturally ambitious announcement at I/O 2026. Demis Hassabis introduced Gemini Omni not as a video generator, but as a world model — a system that builds an internal understanding of reality and reasons about what should happen next inside any given scene. Until now, Google ran a split stack: Veo for video, Imagen for images, separate systems for audio. Omni collapses that into one model that reasons across text, image, video, and audio in the same forward pass.

What's Live Today

Architecture (per the model card)

Gemini Omni Flash is a transformer-based model with native multimodal support for text, vision, video, and audio inputs. Trained on Google's TPUs. Fuses Gemini's reasoning capabilities with the generative capabilities of Veo (video), Nano Banana (image editing), and Project Genie (interactive world simulation) into one architecture.

The model is trained to incorporate real-world understanding of physics, motion, and spatial awareness. Per Google's stage demo, this includes intuitive grasp of gravity, kinetic energy, fluid dynamics, and inter-object spatial relationships. The keynote put it to the test by generating a claymation explainer of protein folding — showing the model moving past pixel-matching to understand actual scientific reality.

Capabilities

• Conversational multi-turn editing. "Swap the background to a city street" → "Make the lighting warmer" → "Stabilize the shot" — each edit is cumulative, not regenerative.
• Reference any combination of inputs. Upload images for characters, audio for pacing, video for motion patterns, text for storyboard direction. The model reasons across all of them in one forward pass.
• Character + voice consistency. Identity and voice preserved across every scene in multi-shot output.
• Custom AI avatars. Face + voice cloning to drop yourself into the action.
• World-model physics. Generated scenes respect gravity, fluid dynamics, kinetic momentum — not perfect, but a marked improvement over pixel-prediction video models.

What Omni Will NOT Do (Deliberately)

Audio editing of existing video is deliberately held back over deepfake risk. You can generate new video with new audio. You can edit visuals on existing video. You cannot replace or modify audio on a video you upload. Google made this explicit. The hold is in response to the SAG-AFTRA / Motion Picture Association / Disney pressure following last year's nudification-tool scandals.

Provenance Hardened by Default

• SynthID — invisible watermark, switched on by default, cannot be disabled. Designed to survive common post-processing (re-encoding, cropping, screenshot). Verifiable via Gemini app, Gemini in Chrome, and Search.
• C2PA Content Credentials — adds explicit provenance metadata. Coming to Search and Chrome in the coming months.
• 50 million+ SynthID verifications globally as of the I/O 2026 keynote.
• 100 billion+ AI-generated images and videos marked with SynthID since launch.

Current Limits Worth Knowing

• Clip duration: 10 seconds per generation in current Omni Flash. Longer formats planned for future Pro release.
• Output modality: Video with audio only for now. Standalone image and audio outputs planned.
• Audio inputs: Voice references supported; other audio types rolling out soon.
• Complete consistency through edits, complex motion, and perfectly accurate text rendering remain explicit Google-acknowledged challenges. Test prompts should exercise these specifically before committing.

Veo Still Exists — Pick the Right Tool

Veo 3.1 was not killed at I/O 2026. It received an upgrade with native audio, shot framing control, and character consistency across multiple clips, plus better complex scene generation. The split now:

• Veo 3.1 — photorealism specialist. 4K resolution, 8-second clips, broadcast-quality realism, native audio.
• Gemini Omni Flash — the multimodal director. Up to 10 seconds, lower resolution than Veo, but cross-modal reasoning and conversational editing.

Pick Veo when you need maximum visual fidelity for a final cut. Pick Omni when you need conversational iteration across mixed inputs.

Supporting Generative Models

The four generative siblings around Omni — each updated at I/O 2026, each tuned for a specific output type.

On-Device + Open Weights

The lane that doesn't run on Google's servers. Released in April 2026 just before I/O, but the Nano 4 production rollout was confirmed during the keynote. Apache 2.0 licensed, commercially usable, runs on consumer hardware.

Gemma 4 Family

What's Unique About Gemma 4

• Hybrid attention. Interleaved sliding-window and full global attention. Final layer always global. Proportional RoPE on global layers with unified keys/values for memory efficiency.
• Per-Layer Embeddings (PLE). The "E" prefix means Effective parameters. Each decoder layer gets its own small embedding for every token. Embedding tables are large but used only for lookups, so effective parameter count is much smaller than total — enabling on-device inference at sizes that would otherwise need a server.
• 140+ languages supported natively.
• Native audio input on E2B and E4B — neither Llama 4 nor Qwen 3.5 offers this at these sizes.
• Apache 2.0 license — commercially permissive, no usage caps, no API spend.

Toolchain Support (day one)

Hugging Face Transformers, vLLM, llama.cpp, MLX (Apple Silicon), LM Studio, Ollama, NVIDIA NIM and NeMo, Unsloth, SGLang, Keras, Docker, Baseten, AMD ROCm.

Gemini Nano 4 — The Production On-Device Path

Built on Gemma 4 architecture but optimized further by Google for production Android deployment. Currently in AICore Developer Preview. Production "later this year." Runs on the latest generation of specialized AI accelerators from Google (Tensor), MediaTek, and Qualcomm Technologies. On other devices, the models initially run on CPU implementation that doesn't represent final production performance.

What's coming during the preview window: tool calling, structured output, system prompts, and thinking mode in the Prompt API. The compatibility promise from Google: code written today for Gemma 4 will automatically work on Gemini Nano 4-enabled devices.

For DDS workloads, the immediate implication: any on-device feature you might want to ship for Pixel + Samsung Galaxy users this fall can be prototyped against Gemma 4 today, with confidence that it'll work on Nano 4 production hardware later. No lock-in, no waiting for API access.

Every Surface Gemini Now Runs On

The big strategic shift at I/O 2026: the model-to-surface matrix collapsed. A year ago, each Google surface had its own AI feature set. Now nearly every consumer and developer surface runs the same underlying Gemini 3.5 Flash, with surface-specific UI and permissions. Google is treating Gemini as the OS-level intelligence layer rather than as a product line.

Developer Surfaces

Consumer + Workspace Surfaces

Strategic observation: the surfaces that are not getting model-level treatment yet are wearables (Wear OS gets Gemini Intelligence as part of the "advanced device" tier but no specific I/O features) and Google Home/Nest (gradual deepening, but no headline updates). Expect both to get more attention at I/O 2027.

Vibe Coding the New Stack

Vibe coding — intent-based engineering where you describe the goal and let the agent handle implementation — went from indie-hacker novelty in March 2026 to Google's official positioning by May. Two demonstrations from the keynote frame the current state of the art.

The "OS in 12 Hours" Stage Demo

Google's headline showcase: Antigravity 2.0 built the core framework of a working operating system in approximately 12 hours. The numbers behind it:

Named Enterprise Builders Already on 3.5 Flash

Google's launch material cited four enterprises running production workloads:

AI Studio Vibe-Coding Stack

The fastest path from idea to deployed prototype right now:

1. Open ai.studio
2. Pick "Build" tab
3. Describe the app in natural language
4. AI Studio scaffolds:
   - React or Kotlin (web or Android)
   - Firebase database + auth + multiplayer
   - npm packages auto-resolved
   - Workspace API access if needed
   - Custom images via Nano Banana inline
5. Preview live in browser (web) or Android Emulator
6. One-click deploy to Cloud Run (web) or
   Play Console test track (Android)
7. Optional: export to Antigravity for deeper work
   — full context, secrets, conversation history carry over

First two apps deploy to Google Cloud free, no credit card. Beyond that, standard Cloud Run pricing applies. For a typical greenfield prototype, this puts a multiplayer-capable production app in roughly an afternoon — what previously took a sprint.

Android Studio Migration Agent

One of the most quietly important launches. Takes a React Native, web framework, or iOS codebase and converts it to native Kotlin + Jetpack Compose. The actual demo prompt from I/O:

"Migrate this React Native codebase to native Kotlin Android.
Preserve all business logic in the data and domain layers.
Use Jetpack Compose for all UI components.
Target API 35.
Maintain the existing Firebase integration.
Flag anything requiring manual review."

Work that previously justified a three-week sprint plus a consultant invoice now completes in hours. Read the flagged-for-review items carefully — the agent will surface ambiguities rather than guess.

WebMCP — The Emerging Open Standard

Announced at I/O 2026 alongside Antigravity 2.0 and the Android CLI: WebMCP, a proposed open web standard for browser-based AI agents. It's been in Early Preview Program since February 10, 2026. The pitch: recast the developer stack around AI agents that can read, write, and ship code through the browser, using a standard protocol any agent vendor (Google, Anthropic, OpenAI) can implement.

If WebMCP succeeds the way MCP succeeded for tools, browser-based agent interoperability will be the default by I/O 2027. Worth tracking even if you're not building on it yet.

Creator Workflows on the New Stack

The other half of the I/O 2026 launch energy went to creators — short-form video makers, brand designers, music producers. The Omni Flash + Flow Agent + Pics + Pomelli + Stitch stack is now the most coherent end-to-end creator toolchain Google has shipped.

Brand Product Video Pipeline (Omni Flash)

A live production-grade workflow for commercial creators, per Google's documentation:

Input Assets:
  - brand-product-shot.png (e.g., product still life)
  - brand-audio-bumper.mp3 (3-second sting)
  - reference-cinematography.mp4 (motion style reference)

Iteration 1 (text + 3 references):
  "Use the product image as the main reference. Create a premium
  10-second product video with a slow dolly-in, warm studio lighting,
  reflective surface detail, subtle ambient sound, and the text
  'Track every move' appearing on the final beat."

Iteration 2 (conversational refinement):
  "Make the lighting warmer."
  → Model preserves all other elements, adjusts only color temperature.

Iteration 3:
  "Stabilize the shot."
  → Reduces camera motion while keeping framing.

Iteration 4:
  "Replace the text with 'Built for everyday wear'."
  → In-place text edit, no regeneration of underlying scene.

The key insight: Omni's world-model architecture retains environment logic across edits. You're not regenerating the video each time — you're refining a persistent scene.

YouTube Shorts Remix (Free for 18+)

The free tier of Omni Flash exposure. Process:

• Select an eligible Short you want to remix.
• Prompt the edit: "Add myself standing next to the host" or "Change the background to a Boston street scene."
• Omni Flash generates the remix preserving the original's pacing and music.
• Custom AI avatar can be created in the YouTube Create app and reused across remixes.

For creators with a built audience, this becomes a duet/stitch mechanic with infinite remix depth. For brands, it's a participatory campaign primitive.

Google Flow + Flow Agent + Flow Tools

Pomelli + Stitch — Brand Design

For brand builders specifically:

• Pomelli — brand content + website design. Builds on-brand creative for campaigns from a brand spec.
• Stitch — real-time design with voice or text steering. Import existing codebase + design files for on-brand consistency. Reflows ideas as you describe them.

For DDS specifically, the Stitch import-codebase capability is interesting: feed it the existing Atelier 2.1.1 theme files and have it generate new sections that match the design language without re-explaining the palette and typography stack.

Full-Stack Automation Patterns

The breadth of automation surfaces is the structural change. Antigravity isn't the only place you can run agents anymore — Workspace, Chrome, Android, and Search all expose agentic patterns now. Pick the pattern that matches the trigger, not the model.

Dev Loop Automation (Antigravity)

• PR review automations — agent reads diff, runs tests, drafts review comments.
• Test generation — coverage-targeted unit + integration tests from existing code.
• Dependency audits — scheduled scan for outdated packages, CVEs, license issues.
• CI gate enforcement via hooks — pre-commit, pre-push, post-test hooks that fail loudly on regression.
• Doc generation — README, API docs, architecture diagrams auto-updated on code changes.
• Code migration — React Native → Kotlin (via Android Studio Migration Agent), web → Compose, legacy → modern frameworks.

Browser-Driven Automation (Antigravity browser subagent + Chrome Auto Browse)

The browser subagent inside Antigravity runs a specialized model different from the main agent, with tools for clicking, scrolling, typing, reading console logs, DOM capture, screenshots, markdown parsing, and video recording. It routes through the Antigravity Chrome extension and is purpose-built for visual verification — the agent sees the UI exactly as a user does.

• Form filling — invoice submission, application forms, government portals.
• Multi-step research/scraping — agent navigates pagination, follows links, structures output.
• E-commerce checkout testing — full purchase flow validation with real card sandboxes.
• Recurring web jobs via Antigravity scheduled tasks — daily competitor pricing scrapes, weekly SEO position checks.
• Visual UI verification — screenshot diff against design spec, regression catch.

Workspace Agent Patterns

• Gmail AI Inbox — intelligent surfacing, AI draft replies, dismissal of stale threads. Reduces inbox triage from hours to minutes for high-volume users.
• Gmail Live (summer) — voice queries to inbox without thread digging.
• Docs Live (summer) — voice-driven doc creation pulling from Gmail/Drive/Chat/web with permission.
• Talk to Keep (summer) — voice brain-dump → organized notes/lists.
• Workspace APIs callable from AI Studio agents — dashboards on Sheets data, tools that organize Drive folders, apps that consume team documents.

Android Agent Patterns (Android 17)

• Gemini Intelligence — OS-level intelligence layer underneath Android, not a separate app.
• AppFunctions — Android apps expose capabilities to agents through a standardized interface. The Android equivalent of MCP for apps.
• Create My Widget — natural-language widget creation OS-wide.
• Power button activates Gemini in any running app.

Chrome Agent Patterns

• Gemini in Chrome Auto Browse — agentic multi-step tasks like booking travel, entering data, scheduling meetings. AI Pro/Ultra US preview.
• Chrome DevTools for agents — new surface specifically targeting autonomous agents.
• Persistent sidebar context — Gemini understands up to 10 open tabs as a single context group.
• WebMCP — emerging open standard for cross-vendor browser agent interoperability.

Search Agent Patterns

• Information agents — 24/7 background research per topic. Summer rollout to AI Pro/Ultra. The Spark equivalent for ongoing topical monitoring.
• Generative UI with Antigravity — Search builds custom layouts, tables, graphs, simulations on the fly. Free for everyone summer.
• Mini apps — custom dashboards/trackers built in Search for ongoing tasks like wedding planning, home moves, projects.

Gemini Spark — The Consumer-Grade Autonomous Agent

Spark is the consumer Antigravity. The architecture:

• Runs on Google Cloud Run VMs — operates with your phone or laptop completely off.
• Built on Gemini 3.5 Flash plus the Antigravity harness.
• Daily Brief integration for proactive morning surfacing.
• Beta AI Ultra US starts week of May 26, 2026.

Roadmap features Google committed to "throughout the summer":

• Text or email Spark directly with task requests.
• Create custom sub-agents under your direction.
• Authorize payments while specifying budget caps and merchant restrictions.

That last one — payment authorization with budget and merchant constraints — is the boundary Google is approaching cautiously. Expect heavy guardrails and confirmation flows in early beta.

Multi-Agent Orchestration Pattern (Antigravity SDK)

The canonical template emerging from early SDK adoption:

// AGENTS.md — define the team
LeadDeveloper:
  role: orchestrator
  routes:
    - design → Designer
    - code → Coder
    - test → Testing
    - docs → Documentation
  model: gemini-3.5-flash

Designer:
  role: research + UI specification
  tools: [browser, code_editor]
  model: gemini-3.5-flash

Coder:
  role: implementation
  tools: [code_editor, terminal]
  model: gemini-3.5-flash

Testing:
  role: unit + regression tests
  tools: [code_editor, terminal]
  model: gemini-3.5-flash

Documentation:
  role: README + API docs maintenance
  tools: [code_editor]
  model: gemini-3.5-flash

Subagent context isolation is the win — the Coder doesn't see what Designer researched, only the spec it produced. Reduces context bloat at scale.

DDS Real-World Applications

Theory is cheap. Here's where the new stack actually changes how Design Delight Studio operates — three concrete patterns you can implement this week.

These three patterns share a common shape: existing manual or expensive workflow becomes a single API call with declarative agent configuration. That's the structural shift in this I/O launch. Capabilities you'd build a whole product around in 2024 are now one config file in 2026.

Pricing + Quotas — The Honest Map

Three layers of pricing changed at I/O 2026: consumer subscriptions restructured, API rates published for 3.5 Flash, and a new compute-based quota system replaced the old daily prompt caps. Read all three before deciding where to spend.

Consumer Subscription Tiers

Gemini API — Per-Token Pricing

API Pricing Modifiers

• Batch mode — 50% off for asynchronous non-urgent workloads
• Context cache reads — 10% of base input price (e.g., $0.15 for 3.5 Flash)
• Context cache storage — $1.00 to $4.50 per 1M tokens per hour, model-dependent
• Search grounding — 5,000 free prompts/month across the Gemini 3.x family, then $14 per 1,000
• Maps grounding — $14 per 1,000 prompts on 3.x family

The Compute-Based Quota System

The structural change that triggered the AI Pro backlash. Google replaced daily prompt caps with compute-based credits. Factors that drain credit faster:

• Prompt complexity (long context, multi-turn, multimodal input)
• Features used (thinking mode at high level, tool calls, search grounding)
• Chat length (cumulative context within a session)

Refresh: every 5 hours. Hard ceiling: weekly. When you hit the cap on the biggest models, the system shifts you to smaller models — no service interruption, but possible silent quality degradation. Pay-as-you-go top-up credits are available for Antigravity and Flow now. Gemini app top-ups coming.

Regional Pricing Gap

Non-global API regions cost 10% more across the board. If your workload tolerates global region routing, you save $0.15 input and $0.90 output per million tokens. For high-volume agent workloads, that adds up — review your Cloud project's region settings before committing to the new pricing.

Launch-Week Bugs and Workarounds

Five days in, the community has documented seven distinct issues. Five are workable. Two are silent risks that require explicit audits before flipping production. Read every section before subscribing or migrating.

cd "$env:LOCALAPPDATA\Programs\Antigravity\resources"
Rename-Item app.asar app.asar.bak

cd ~/Antigravity-x64
sudo chown root:root chrome-sandbox
sudo chmod 4755 chrome-sandbox

The Roadmap — Dated and Committed

Every dated milestone Google committed to during the I/O 2026 keynote and accompanying blog posts. Use this for sprint planning.

Late May 2026 (this week and next)

June 2026

Summer 2026 ("this summer")

• Information agents in Search → AI Pro/Ultra subscribers
• Generative UI with Antigravity in Search → free for everyone
• Universal Cart across Search + Gemini app, then YouTube + Gmail
• Daily Brief expansion beyond US
• Gmail Live, Docs Live, Talk to Keep → AI Pro/Ultra
• Google Pics GA → AI Pro/Ultra + Workspace business preview
• Gemini Spark feature drops throughout
• Mini apps in Search → subscribers

Fall 2026

• Android XR audio glasses launch — Gentle Monster, Warby Parker, Samsung
• Android Halo — works with Gemini Spark
• Gemini Nano 4 production rollout

"Coming Months" (no firm date)

• Antigravity in Gemini Enterprise for existing customers
• C2PA Content Credentials verification in Search + Chrome
• Vertex AI public Omni Flash API
• Spark text/email direct interaction
• Spark custom sub-agents
• Spark payment authorization with budget/merchant constraints

Regional Gaps to Watch

Many of the new Workspace features launch US-only. Daily Brief is US only at launch. Spark beta is US only. Ask YouTube is US English only initially. Non-global API regions cost 10% more. Build expansion plans assuming US-first rollout with global catch-up over 3-6 months.

Frequently Asked Questions

Continue Your Training

The three Academy masterclasses most directly connected to this material. Read them in order if you're new to the agent stack.