@oriro/orirocli 0.1.9 → 0.1.11

package/skills/craft/google-ai-latest/SKILL.md

@@ -1,682 +1,682 @@
----
-watermark: ORIRO
-name: google-ai-latest
-provider: ORIRO.ai
-copyright: Copyright (c) 2026 ORIRO.ai
-description: >
-  Google AI 2026 — Gemma 4 family, Gemini 3 family, Vertex/AI Studio pricing
-  matrix. Activate for ANY task involving: Gemma 4 (E2B/E4B/26B/31B), Gemini
-  3 Pro/Flash/Flash-Lite, Gemini 3.1, model selection for ORIRO.ai agents,
-  D1 google_ai_capabilities table updates, on-device AI (Gemini Nano/Chrome),
-  Imagen 4, Veo 3.1, embedding models, pricing decisions, free tier limits,
-  Vertex AI vs Google AI Studio selection, or any question about which Google
-  model to use for what ORIRO.ai task.
-  Sources: deepmind.google/models/gemma/gemma-4, ai.google.dev/gemma/docs,
-  ai.google.dev/gemini-api/docs/pricing, cloud.google.com Vertex AI pricing.
-  Last read: June 15, 2026.
----
-
-# Google AI 2026 — Complete Model & Pricing Reference
-
-**Sources read live June 15, 2026:**
-
-- https://deepmind.google/models/gemma/gemma-4/
-- https://cloud.google.com/blog/products/ai-machine-learning/gemma-4-available-on-google-cloud
-- https://ai.google.dev/gemma/docs/releases
-- https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/3-pro
-- https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models
-- https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing
-
----
-
-## 1. GEMMA 4 — The Game-Changer Open Model
-
-**Released:** April 2, 2026 — now **GA on Google Cloud** (Vertex AI Model
-Garden fully-managed/serverless, Cloud Run on NVIDIA RTX PRO 6000 Blackwell,
-GKE, Cloud TPUs, Sovereign Cloud)
-**License:** Apache 2.0 (first time in Gemma line — truly open)
-**Architecture:** Dense + Mixture-of-Experts (MoE) variants
-**Built from:** Gemini 3 research and technology
-
-### Gemma 4 Model Sizes — Exact Specs
-
-| Model       | Params (total) | Params (active) | Context | Architecture | Best for                                   |
-| ----------- | -------------- | --------------- | ------- | ------------ | ------------------------------------------ |
-| E2B         | 5.1B           | 2.3B            | 128K    | MoE + PLE    | Phones, IoT, edge                          |
-| E4B         | 8.0B           | 4.5B            | 128K    | MoE + PLE    | Laptops, Raspberry Pi                      |
-| 12B Unified | 12B            | 12B (dense)     | 256K    | Dense        | Mid GPU; text+audio+image (new Jun 3 2026) |
-| 26B A4B     | 25.2B          | 3.8B            | 256K    | MoE          | Consumer GPU (RTX 4090)                    |
-| 31B         | 30.7B          | 30.7B (dense)   | 256K    | Dense        | Workstations, H100                         |
-
-**New since launch:** Gemma 4 **12B Unified** (released Jun 3, 2026 — text,
-audio, image input, 256K context). Gemma 4 **MTP** (Multi-Token Prediction,
-Apr 16) variants exist for E2B/E4B/31B/26B A4B for faster decoding.
-
-**PLE = Per-Layer Embeddings:** Edge models activate far fewer parameters
-than total count. The 26B MoE has speed comparable to a 4B model.
-
-### Gemma 4 Capabilities
-
-```
-Multimodal (ALL models):
-  - Text input → Text output ✅
-  - Image input (variable resolution, variable aspect ratio) ✅
-  - Video input ✅ (all models)
-  - Audio input ✅ (E2B + E4B only natively)
-
-Reasoning: Built-in thinking mode (configurable — can enable/disable)
-Context: 128K (edge) / 256K (medium/large)
-Languages: 140+ languages
-Function calling: Native, built-in — critical for agentic workflows
-Coding: AIME 2026 = 89.2%, Codeforces ELO = 2150
-MMLU Pro: 85.2% (31B)
-Arena AI rank: #3 open model globally (31B), #6 (26B)
-```
-
-### Gemma 4 Pricing
-
-```
-Open weights = FREE to download and self-host
-API hosting:
-  Gemma 4 31B (OpenRouter free tier): $0/M tokens ← use this
-  Gemma 4 26B (Vertex AI managed): $0.15/M input, $0.60/M output,
-    cache hit $0.015/M  ← updated Jun 2026 (was $0.06/$0.30)
-  Third-party hosted: $0.15–0.60/M tokens typically
-
-Self-hosting cost:
-  26B quantized (RTX 4090): ~$0.001-0.005/M tokens
-  31B full bfloat16: needs H100 80GB or equivalent
-  26B MoE: 3.8B active → runs fast on consumer GPU
-
-Gemma 4 vs proprietary:
-  Claude Sonnet: ~12-25x more expensive per token
-  GPT-4o: ~5-10x more expensive per token
-  Gemini 3 Flash: direct competitor (same quality tier)
-```
-
-### Gemma 4 Architecture Innovations
-
-```
-1. Mixture-of-Experts (MoE):
-   26B total params, only 3.8B activated per inference
-   = 4B model speed at 26B model quality
-
-2. Per-Layer Embeddings (PLE) in edge models:
-   Maximize parameter efficiency
-   E2B: 5.1B params, behaves like ~2.3B
-
-3. Visual token budget (configurable):
-   Token budgets: 70, 140, 280, 560, 1120
-   Higher budget = more detail, more compute
-   Lower budget = faster inference for simple visual tasks
-   Tip: use 280 for most ORIRO.ai tasks, 560 for complex image analysis
-
-4. KV cache sharing with assistants:
-   Avoids recomputing context in multi-turn conversations
-   Critical for ORIRO.ai agent chains
-
-5. Thinking mode (configurable):
-   Can enable step-by-step reasoning or skip for speed
-   Enable for complex tasks, disable for Flash-level speed
-```
-
----
-
-## 2. GEMINI 3 FAMILY — Current Production Models
-
-### Gemini 3 Model Matrix (May 2026)
-
-| Model                  | Input Price/1M      | Output Price/1M         | Context | Status                         |
-| ---------------------- | ------------------- | ----------------------- | ------- | ------------------------------ |
-| Gemini 3.1 Pro         | $2.00 (>200K $4.00) | $12.00 (>200K $18.00)   | 1M      | GA                             |
-| Gemini 3.5 Flash       | $1.50               | $9.00                   | 1M      | GA (new in matrix Jun 2026)    |
-| Gemini 3.1 Flash-Lite  | $0.25               | $1.50                   | 1M      | GA (repriced from $0.10/$0.40) |
-| Gemini 3.1 Flash Image | $0.50               | $3.00 +$60/1M img out   | 1M      | GA                             |
-| Gemini 3 Pro Image     | $2.00               | $12.00 +$120/1M img out | 1M      | GA (new)                       |
-| Gemini 3 Flash         | $0.50 (audio $1)    | $3.00                   | 1M      | Preview                        |
-| Gemini 3.1 Flash       | $0.30               | $2.50                   | 1M      | Superseded by 3.5 Flash        |
-| Gemini 2.5 / 2.0 (all) | —                   | —                       | 1M      | Retired (after June 1)         |
-
-**CRITICAL:** Gemini 2.0 Flash, 2.0 Flash-Lite, and 2.5 Flash retired after
-June 1, 2026 — migrate any remaining calls to Gemini 3.5 Flash (standard) or
-3.1 Flash-Lite (budget). Note 3.5 Flash now costs $1.50/$9.00 — ~5x the old
-3.1 Flash; for cost-sensitive high volume, prefer 3.1 Flash-Lite ($0.25/$1.50)
-or Gemma 4 26B ($0.15/$0.60).
-
-### Gemini 3.1 Pro — What It Excels At
-
-```
-ARC-AGI-2 score: 77.1% (2x its predecessor)
-Best for: Complex reasoning, long documents, advanced code generation
-Pricing: $2.00/$12.00 per 1M tokens (>200K prompt: $4.00/$18.00)
-Context: 1M tokens (industry-leading for proprietary model)
-Use in ORIRO.ai: Strategy Builder, complex Campaign Exploder logic
-```
-
-### Gemini 3.1 Flash — The Sweet Spot
-
-```
-Best price:performance ratio for ORIRO.ai agents
-$0.30/$2.50 per 1M tokens (12-25x cheaper than Claude Sonnet)
-Speed: Faster than Pro, excellent quality
-Use in ORIRO.ai: All Flash Prompts, Brand Voice, Email Campaign
-Batch mode: 50% discount → $0.15/$1.25 for async workloads
-```
-
-### Gemini 3.1 Flash-Lite — Budget Champion
-
-```
-$0.10/$0.40 per 1M tokens — cheapest production-grade model
-GA since May 7, 2026
-Use in ORIRO.ai: High-volume content generation, embeddings
-No OpenAI equivalent at this price point
-```
-
----
-
-## 3. SPECIALIZED MODELS — Image, Video, Audio
-
-### Imagen 4 (Current as of May 2026)
-
-```
-Models: imagen-4.0-generate-001, imagen-4.0-ultra-generate-001,
-        imagen-4.0-fast-generate-001
-Status: Preview (paid tier only)
-Pricing:
-  512px output: $0.045/image (747 tokens)
-  1K output:    $0.067/image (1120 tokens)
-  2K output:    $0.101/image (1680 tokens)
-  4K output:    $0.151/image (2520 tokens)
-Use in ORIRO.ai: Agent avatar generation, campaign image creation
-Notable: "significantly better text rendering" vs Imagen 3
-```
-
-### Veo 3.1 (Current as of May 2026)
-
-```
-Models: veo-3.1-generate-preview, veo-3.1-fast-generate-preview,
-        veo-3.1-lite-generate-preview
-Status: Preview (paid tier only for API)
-Pricing: 10 min of 1080p = ~$240
-Use in ORIRO.ai: Phase 4 video pipeline (Veo-powered Flash agents)
-D1 table: active model = veo-3.1 (confirmed)
-Note: Only charged if video successfully generates
-      Audio processing issues can prevent generation (no charge)
-```
-
-### Embedding Models
-
-```
-text-embedding-005: Current production embedding model
-  Use in ORIRO.ai: Agent knowledge base, semantic search
-  Free tier: available
-  Pricing: low (not published as of May 2026)
-
-multimodal-embedding-001: First multimodal embedding
-  Maps text, images, video, audio, PDFs → unified embedding space
-  New May 2026 — evaluate for future ORIRO.ai agent memory
-```
-
----
-
-## 4. FREE TIER LIMITS (May 2026 — verified)
-
-**IMPORTANT:** Free tier limits changed significantly in 2026.
-Google cut quotas 50-80% in December 2025, more cuts April 1, 2026.
-
-```
-Google AI Studio (browser interface): Always free, no limits in UI
-API free tier (as of May 2026):
-  Flash models: higher RPM/TPM after May 7 2026 GA rollout
-  Pro models: 50 requests/day (cut from 100 on April 1)
-  All models: full 1M token context window
-  Data policy: FREE tier data may be used for model training
-               PAID tier: your data is NOT used for training
-
-Rate limit dimensions:
-  RPM = Requests Per Minute
-  TPM = Tokens Per Minute
-  RPD = Requests Per Day
-  Note: Check actual limits in AI Studio — numbers change monthly
-
-Free tier strategy for ORIRO.ai:
-  Development/testing: Free tier is sufficient
-  Production: Enable billing (spend caps prevent surprises)
-  Private user data: MUST use paid tier (data not used for training)
-```
-
-### Gemma 4 Free Access
-
-```
-OpenRouter free tier: Gemma 4 31B = $0/M tokens (rate limited)
-Google AI Studio:     Gemma 4 31B + 26B free for development
-Third-party hosted:   Replicate, Together AI, etc. — check current rates
-
-Self-hosting:
-  E4B: runs on modern laptop, phone, Raspberry Pi
-  26B MoE: RTX 4090 or Apple Silicon (24GB+ unified memory)
-  31B dense: H100 80GB (not for self-hosting on consumer hardware)
-```
-
----
-
-## 5. VERTEX AI vs GOOGLE AI STUDIO
-
-```
-                    Google AI Studio        Vertex AI
-Development:        ✅ Instant access        Requires project setup
-Pricing:            Free tier + per-token   Per-token, enterprise SLAs
-Data privacy:       Free=training data      Private, GDPR compliant
-Production:         Small-medium workloads  Enterprise workloads
-Rate limits:        Lower (free tier)       Higher (paid)
-Non-global endpoints: N/A                  Pricing from July 1, 2026
-Use for ORIRO.ai:   MVP + development       Post-scale (phase 5+)
-
-ORIRO.ai decision: Google AI Studio API for now
-                   Vertex AI when user data privacy becomes critical
-                   or when rate limits become bottleneck
-```
-
----
-
-## 6. ORIRO.ai — SPECIFIC MODEL DECISIONS
-
-### Current D1 google_ai_capabilities Table (Verified May 2026)
-
-```sql
--- Active models in production:
-flash_primary:  gemini-2.5-flash    ← UPDATE to gemini-3.1-flash
-learn_primary:  gemini-2.5-flash    ← UPDATE to gemini-3.1-flash
-video:          veo-3.1             ← CORRECT ✅
-image:          gemini-3.1-flash-image  ← CORRECT ✅
-embedding:      text-embedding-005  ← CORRECT ✅
-
--- Deprecated (update before June 1):
-gemini-2.0-Flash-Lite    → migrate to gemini-3.1-flash-lite
-gemini-2.5-flash         → migrate to gemini-3.1-flash
-```
-
-### Flash Prompts 1-8 — Right Model for Each
-
-```
-Flash Prompt 1 (Campaign Exploder):   gemini-3.1-flash ← $0.30/M
-Flash Prompt 2 (Brand Voice):         gemini-3.1-flash
-Flash Prompt 3 (Strategy Builder):    gemini-3.1-pro   ← complex
-Flash Prompt 4 (Content Calendar):    gemini-3.1-flash
-Flash Prompt 5 (Ad Copy):             gemini-3.1-flash
-Flash Prompt 6 (Email Campaign):      gemini-3.1-flash
-Flash Prompt 7 (Social Content):      gemini-3.1-flash-lite ← cheapest
-Flash Prompt 8 (Flash Dashboard):     gemini-3.1-flash
-```
-
-### Tier Model Integration
-
-```
-T1 (Free — Chrome Gemini Nano):
-  Uses on-device Chrome extension
-  Gemma 4 E2B equivalent performance
-  ORIRO pays $0 compute ✅
-  Limit: text only, no image generation
-
-T2 (Google Pro $0.99/mo):
-  Access to gemini-3.1-flash via user's Google API key
-  ORIRO pays $0 compute ✅
-  User brings their own key
-  Enable image generation via Imagen 4
-
-T3 (BYO Any Key $0.99/mo):
-  User supplies Gemini/OpenAI/Anthropic key
-  ORIRO passes key through — pays $0
-  Unlock Veo 3.1 for video agents
-
-ORIRO Cost Model: $0 compute on all tiers ✅ (confirmed correct)
-```
-
-### Gemma 4 Opportunity for ORIRO.ai
-
-```
-PROBLEM TO SOLVE: T1 users get limited Gemini Nano (Chrome only)
-OPPORTUNITY: Gemma 4 E4B runs on ANY device (no Chrome required)
-
-Option A: Offer Gemma 4 local mode for T1 users
-  - User downloads E4B (8B params) to their machine
-  - ORIRO agents run locally — no API calls, no cost
-  - 128K context, multimodal, function calling
-  - Privacy: all data stays on device
-
-Option B: Use Gemma 4 26B via OpenRouter free tier
-  - Free API access for development/low-volume
-  - $0.06/M for production
-  - Better than Nano for complex agent tasks
-
-Option C: Fine-tune Gemma 4 E4B on ORIRO agent personas
-  - Apache 2.0 = can fine-tune and redistribute
-  - One-time cost: $50-500 for LoRA fine-tune
-  - Inference cost after: same as base ($0)
-  - Create "ORIRO-Agent" specialized model
-
-RECOMMENDATION: Implement Option B for T1 power users
-Add Gemma 4 26B as T1 premium option alongside Gemini Nano
-```
-
-### Agent Memory Architecture (Gemma 4 Advantage)
-
-```
-Current ORIRO: gemini-2.5-flash for learn_primary
-Better with Gemma 4 31B reasoning:
-  256K context = full campaign history in one prompt
-  Native function calling = structured agent tool use
-  Thinking mode = step-by-step campaign strategy
-  Cost: $0.06/M vs $0.30/M for Flash (5x cheaper for reasoning)
-
-For Knowledge Base agents:
-  text-embedding-005 → embed user content
-  multimodal-embedding-001 → embed images+text together
-  Store in ORIRO's D1 or Cloudflare Vectorize
-```
-
----
-
-## 7. MODEL SELECTION DECISION TREE
-
-```
-Is it a REASONING task (strategy, analysis, complex copy)?
-  → gemini-3.1-pro ($2.00/M) or Gemma 4 31B ($0.06/M)
-
-Is it a STANDARD agent task (content, email, social)?
-  → gemini-3.1-flash ($0.30/M)
-
-Is it HIGH-VOLUME, SIMPLE text generation?
-  → gemini-3.1-flash-lite ($0.10/M) or Gemma 4 26B ($0.06/M)
-
-Is it IMAGE GENERATION?
-  → imagen-4.0-fast ($0.045/img 512px) for speed
-  → imagen-4.0-ultra for quality
-
-Is it VIDEO GENERATION?
-  → veo-3.1 (paid tier, ~$240/10min 1080p)
-
-Is it EMBEDDING?
-  → text-embedding-005 (text only)
-  → multimodal-embedding-001 (text+image+video+audio)
-
-Is it ON-DEVICE (T1 Chrome user)?
-  → Gemini Nano (current)
-  → Gemma 4 E4B (future — better capability)
-
-Does user have their own API key (T2/T3)?
-  → Pass through their key, ORIRO pays $0
-```
-
----
-
-## 8. CRITICAL MIGRATION — DO BEFORE JUNE 1, 2026
-
-```
-Deprecated models to replace:
-  gemini-2.0-Flash     → gemini-3.1-flash
-  gemini-2.0-Flash-Lite → gemini-3.1-flash-lite
-  gemini-2.5-flash     → gemini-3.1-flash (already mostly migrated)
-
-Files to update in ORIRO.ai:
-  D1 migration: infra/cloudflare/d1/migrations/
-  Update google_ai_capabilities table
-  Update any hardcoded model strings in apps/web
-
-Files to update in <project>-intel:
-  Update summarizer model string from old Gemini to gemini-3.1-flash
-  <project>-intel-src/main.py → grep for "gemini" model strings
-  <project>-intel-src/config.py → GEMINI_MODEL constant
-
-Files to update in the app router:
-  Update routes that reference old Gemini model strings
-```
-
----
-
-## 9. BENCHMARKS — WHERE GEMMA 4 WINS
-
-```
-Gemma 4 31B vs competition:
-  MMLU Pro:        85.2% (better than most 70B models)
-  AIME 2026:       89.2% math reasoning
-  Codeforces ELO:  2150 (top-tier coding)
-  Arena AI rank:   #3 globally among open models
-  Arena AI score:  1452 (text only)
-
-Gemma 4 26B MoE (only 3.8B active):
-  Arena AI rank:   #6 globally
-  Arena AI score:  1441
-  Speed:           Comparable to 4B model
-
-This means ORIRO.ai agents can use Gemma 4 26B and get:
-  - Near-frontier reasoning quality
-  - 4B model speed
-  - 256K context window
-  - $0.06/M tokens (5x cheaper than Gemini 3.1 Flash)
-```
-
----
-
-## 10. QUICK REFERENCE — MODEL STRINGS
-
-```python
-# Current production model strings (May 2026):
-GEMINI_PRO         = "gemini-3.1-pro"
-GEMINI_FLASH       = "gemini-3.1-flash"
-GEMINI_FLASH_LITE  = "gemini-3.1-flash-lite"
-GEMINI_FLASH_IMAGE = "gemini-3.1-flash-image"
-GEMINI_EMBED       = "text-embedding-005"
-IMAGEN             = "imagen-4.0-generate-001"
-IMAGEN_FAST        = "imagen-4.0-fast-generate-001"
-VEO                = "veo-3.1-generate-preview"
-GEMMA_4_31B        = "gemma-4-31b-it"         # OpenRouter free
-GEMMA_4_26B        = "gemma-4-27b-it"         # Google API
-
-# Old strings to REPLACE (deprecated June 1):
-OLD_FLASH_LITE = "gemini-2.0-flash-lite"    # ← REMOVE
-OLD_FLASH      = "gemini-2.0-flash"         # ← REMOVE
-OLD_25_FLASH   = "gemini-2.5-flash"         # ← UPDATE to 3.1-flash
-
-# <project>-intel summarizer update:
-APP_SUMMARIZER   = "gemini-3.1-flash"
-APP_SCORER       = "claude-haiku-4-5-20251001"  # unchanged
-```
-
----
-
-## 11. GOOGLE I/O MAY 19-20 2026 — WHAT TO WATCH
-
-```
-Expected announcements (monitor and update ORIRO D1 + <project>):
-  - Gemini 3.2 Flash (rumored) → update flash_primary if released
-  - Gemini 3 Flash GA (been preview since Dec 2025)
-  - "Deep Think" mode rollout (advanced reasoning)
-  - Veo 4 (next video generation model)
-  - Imagen 4 updates
-  - Free tier rate limit changes
-
-Update procedure:
-  1. Check ai.google.dev/gemini-api/docs/pricing
-  2. Update D1 google_ai_capabilities table via migration
-  3. Update <project>-intel model strings
-  4. Update the app router routes
-  5. Test all Flash Prompts with new model strings
-```
-
----
-
-_Read live from Google DeepMind, Google AI Developers, Vertex AI docs_
-_Prices and models verified May 19, 2026_
-_Next scheduled refresh: Monday May 25, 2026 (ClaudeSkills-GoogleAI-Refresh task)_
-
----
-
-## 12. GOOGLE I/O 2026 — LIVE ANNOUNCEMENTS (May 19, 2026)
-
-**Source:** Tom's Guide live blog + BusinessToday + TechRadar — read May 19, 2026
-
-### NEW MODELS ANNOUNCED TODAY
-
-#### Gemini 3.5 Flash — JUST RELEASED
-
-```
-Status:    LIVE (announced keynote May 19, 2026)
-Described: "Strongest agentic and coding model"
-Benchmarks: Surpasses Gemini 3.1 Pro on:
-  - Coding
-  - Reasoning
-  - Multimodal understanding
-  - Terminal-Bench 2.1
-Use case:  Developers, businesses, complex AI tasks
-Speed:     Faster and cheaper than other advanced models
-Capabilities:
-  - Plan and organise tasks
-  - Build and improve solutions iteratively
-  - Handle practical business and software work
-  - Develop new apps
-  - Maintain/update codebases
-  - Prepare financial documents
-
-ORIRO.ai action: UPDATE flash_primary to gemini-3.5-flash
-<project> action:    UPDATE summarizer to gemini-3.5-flash
-Model string:    "gemini-3.5-flash" (confirm in AI Studio)
-```
-
-#### Gemini 3.5 Pro — COMING NEXT MONTH (June 2026)
-
-```
-Status:    Announced, not yet released
-Described: "Powerful version of Flash"
-ETA:       June 2026
-Capabilities: Across all Google products and platforms
-ORIRO action:  Plan migration from gemini-3.1-pro → gemini-3.5-pro
-               Update D1 when released
-```
-
-#### Gemini Spark — NEW AGENTIC AI
-
-```
-Status:    LIVE for AI Ultra users (US only initially)
-Tier:      AI Ultra ($100/month — new tier announced today)
-Built on:  Gemini 3.5 + Antigravity harness
-Type:      "24/7 personal AI agent"
-Connects:  Gmail, Docs, Slides, Google Workspace
-Functions:
-  - Declutter inboxes (analyze + summarize emails)
-  - Auto-archive newsletters
-  - Meeting summaries
-  - Structure info from Google apps
-  - Personal news briefings (follows stories over time)
-
-ORIRO.ai relevance:
-  - Gemini Spark = the agentic model ORIRO agents should aspire to
-  - Antigravity harness = Google's agent orchestration layer
-  - Study Antigravity 2.0 architecture for ORIRO agent design
-  - Flash Agents (ORIRO) + Gemini 3.5 Flash = competitive offering
-```
-
-#### Gemini Omni — VIDEO EDITING MODEL
-
-```
-Status:    Announced at I/O 2026
-Type:      Video generation + editing within Gemini app
-Powered by: Veo 3.1 (confirmed)
-Capability: Generate AND edit videos directly in Gemini chat
-ORIRO action: Monitor API availability for Veo phase (Phase 4)
-```
-
-#### Antigravity 2.0 — AGENTIC CODING SYSTEM
-
-```
-Status:    Showcased at I/O 2026
-Achievement: Created core components of an OS in 12 hours
-  - 93 separate sub-agents deployed
-  - 2.6 billion tokens generated
-  - Core OS framework in ~12 hours
-
-ORIRO relevance: This is the architecture for multi-agent coordination
-  Antigravity = orchestrator launching sub-agents
-  ORIRO's Flash Agents = similar sub-agent pattern
-  Study Antigravity for ORIRO's Phase 6 (1000+ agents)
-```
-
-### UPDATED MODEL PRIORITY FOR ORIRO.ai
-
-```
-IMMEDIATE UPDATES NEEDED (post I/O):
-
-1. D1 google_ai_capabilities — update today:
-   flash_primary:  gemini-3.1-flash → gemini-3.5-flash
-   learn_primary:  gemini-3.1-flash → gemini-3.5-flash
-
-2. <project>-intel:
-   APP_SUMMARIZER: gemini-3.1-flash → gemini-3.5-flash
-
-3. the app router routes:
-   Update any Gemini model string references
-
-JUNE 2026 (when Gemini 3.5 Pro releases):
-   flash_primary strategy: gemini-3.5-pro for complex agents
-   gemini-3.5-flash for standard agents (cost-optimized)
-
-NEW MODEL STRINGS (I/O 2026):
-   GEMINI_35_FLASH = "gemini-3.5-flash"    ← NEW ✅
-   GEMINI_35_PRO   = "gemini-3.5-pro"      ← COMING JUNE
-   GEMINI_SPARK    = AI Ultra tier only
-   GEMINI_OMNI     = Veo integration (monitor)
-```
-
-### SMART GLASSES — HARDWARE CONTEXT
-
-```
-Announced: Samsung 'Intelligent Eyewear' — fall 2026
-Also: Xreal Project Aura, other Android XR partners
-Powered by: Gemini (on-device via Gemini Nano equivalent)
-
-ORIRO opportunity:
-  Android XR glasses = new platform for ORIRO agents
-  Phase 6+: ORIRO agents accessible via smart glasses
-  Agent commands via voice → Gemini → ORIRO API
-```
-
-### I/O 2026 DEVELOPER KEYNOTE (May 20 — Tomorrow)
-
-```
-Session to watch: "What's new in the Gemma open model family"
-  - New Gemma additions will be revealed
-  - Practical deployment tools at scale
-  - Update this skill after May 20 developer keynote
-
-Expected from developer keynote:
-  - Gemma 4 updates or new Gemma 4 variants
-  - Vertex AI agent platform updates
-  - Google AI Studio new features
-  - Developer tools for Antigravity/agent building
-```
-
-### UPDATED QUICK REFERENCE (Post I/O 2026)
-
-```python
-# CURRENT production model strings (May 19, 2026 post-I/O):
-GEMINI_35_FLASH    = "gemini-3.5-flash"     # NEW — use this now
-GEMINI_31_PRO      = "gemini-3.1-pro"       # Until 3.5 Pro releases
-GEMINI_FLASH_LITE  = "gemini-3.1-flash-lite"
-GEMINI_FLASH_IMAGE = "gemini-3.1-flash-image"
-GEMINI_EMBED       = "text-embedding-005"
-IMAGEN             = "imagen-4.0-generate-001"
-VEO                = "veo-3.1-generate-preview"  # Omni coming
-GEMMA_4_31B        = "gemma-4-31b-it"
-GEMMA_4_26B        = "gemma-4-27b-it"
-
-# Deprecated (remove before June 1):
-OLD = ["gemini-2.0-flash", "gemini-2.0-flash-lite", "gemini-2.5-flash"]
-
-# ORIRO.ai D1 update SQL:
-# UPDATE google_ai_capabilities
-# SET model_id = 'gemini-3.5-flash'
-# WHERE capability_key IN ('flash_primary', 'learn_primary');
-```
-
-_I/O 2026 Day 1 keynote: May 19, 2026 — read live from Tom's Guide + BusinessToday_
-_Developer keynote Day 2: May 20, 2026 — UPDATE after Gemma session_
+---
+watermark: ORIRO
+name: google-ai-latest
+provider: ORIRO.ai
+copyright: Copyright (c) 2026 ORIRO.ai
+description: >
+  Google AI 2026 — Gemma 4 family, Gemini 3 family, Vertex/AI Studio pricing
+  matrix. Activate for ANY task involving: Gemma 4 (E2B/E4B/26B/31B), Gemini
+  3 Pro/Flash/Flash-Lite, Gemini 3.1, model selection for ORIRO.ai agents,
+  D1 google_ai_capabilities table updates, on-device AI (Gemini Nano/Chrome),
+  Imagen 4, Veo 3.1, embedding models, pricing decisions, free tier limits,
+  Vertex AI vs Google AI Studio selection, or any question about which Google
+  model to use for what ORIRO.ai task.
+  Sources: deepmind.google/models/gemma/gemma-4, ai.google.dev/gemma/docs,
+  ai.google.dev/gemini-api/docs/pricing, cloud.google.com Vertex AI pricing.
+  Last read: June 15, 2026.
+---
+
+# Google AI 2026 — Complete Model & Pricing Reference
+
+**Sources read live June 15, 2026:**
+
+- https://deepmind.google/models/gemma/gemma-4/
+- https://cloud.google.com/blog/products/ai-machine-learning/gemma-4-available-on-google-cloud
+- https://ai.google.dev/gemma/docs/releases
+- https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/3-pro
+- https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models
+- https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing
+
+---
+
+## 1. GEMMA 4 — The Game-Changer Open Model
+
+**Released:** April 2, 2026 — now **GA on Google Cloud** (Vertex AI Model
+Garden fully-managed/serverless, Cloud Run on NVIDIA RTX PRO 6000 Blackwell,
+GKE, Cloud TPUs, Sovereign Cloud)
+**License:** Apache 2.0 (first time in Gemma line — truly open)
+**Architecture:** Dense + Mixture-of-Experts (MoE) variants
+**Built from:** Gemini 3 research and technology
+
+### Gemma 4 Model Sizes — Exact Specs
+
+| Model       | Params (total) | Params (active) | Context | Architecture | Best for                                   |
+| ----------- | -------------- | --------------- | ------- | ------------ | ------------------------------------------ |
+| E2B         | 5.1B           | 2.3B            | 128K    | MoE + PLE    | Phones, IoT, edge                          |
+| E4B         | 8.0B           | 4.5B            | 128K    | MoE + PLE    | Laptops, Raspberry Pi                      |
+| 12B Unified | 12B            | 12B (dense)     | 256K    | Dense        | Mid GPU; text+audio+image (new Jun 3 2026) |
+| 26B A4B     | 25.2B          | 3.8B            | 256K    | MoE          | Consumer GPU (RTX 4090)                    |
+| 31B         | 30.7B          | 30.7B (dense)   | 256K    | Dense        | Workstations, H100                         |
+
+**New since launch:** Gemma 4 **12B Unified** (released Jun 3, 2026 — text,
+audio, image input, 256K context). Gemma 4 **MTP** (Multi-Token Prediction,
+Apr 16) variants exist for E2B/E4B/31B/26B A4B for faster decoding.
+
+**PLE = Per-Layer Embeddings:** Edge models activate far fewer parameters
+than total count. The 26B MoE has speed comparable to a 4B model.
+
+### Gemma 4 Capabilities
+
+```
+Multimodal (ALL models):
+  - Text input → Text output ✅
+  - Image input (variable resolution, variable aspect ratio) ✅
+  - Video input ✅ (all models)
+  - Audio input ✅ (E2B + E4B only natively)
+
+Reasoning: Built-in thinking mode (configurable — can enable/disable)
+Context: 128K (edge) / 256K (medium/large)
+Languages: 140+ languages
+Function calling: Native, built-in — critical for agentic workflows
+Coding: AIME 2026 = 89.2%, Codeforces ELO = 2150
+MMLU Pro: 85.2% (31B)
+Arena AI rank: #3 open model globally (31B), #6 (26B)
+```
+
+### Gemma 4 Pricing
+
+```
+Open weights = FREE to download and self-host
+API hosting:
+  Gemma 4 31B (OpenRouter free tier): $0/M tokens ← use this
+  Gemma 4 26B (Vertex AI managed): $0.15/M input, $0.60/M output,
+    cache hit $0.015/M  ← updated Jun 2026 (was $0.06/$0.30)
+  Third-party hosted: $0.15–0.60/M tokens typically
+
+Self-hosting cost:
+  26B quantized (RTX 4090): ~$0.001-0.005/M tokens
+  31B full bfloat16: needs H100 80GB or equivalent
+  26B MoE: 3.8B active → runs fast on consumer GPU
+
+Gemma 4 vs proprietary:
+  Claude Sonnet: ~12-25x more expensive per token
+  GPT-4o: ~5-10x more expensive per token
+  Gemini 3 Flash: direct competitor (same quality tier)
+```
+
+### Gemma 4 Architecture Innovations
+
+```
+1. Mixture-of-Experts (MoE):
+   26B total params, only 3.8B activated per inference
+   = 4B model speed at 26B model quality
+
+2. Per-Layer Embeddings (PLE) in edge models:
+   Maximize parameter efficiency
+   E2B: 5.1B params, behaves like ~2.3B
+
+3. Visual token budget (configurable):
+   Token budgets: 70, 140, 280, 560, 1120
+   Higher budget = more detail, more compute
+   Lower budget = faster inference for simple visual tasks
+   Tip: use 280 for most ORIRO.ai tasks, 560 for complex image analysis
+
+4. KV cache sharing with assistants:
+   Avoids recomputing context in multi-turn conversations
+   Critical for ORIRO.ai agent chains
+
+5. Thinking mode (configurable):
+   Can enable step-by-step reasoning or skip for speed
+   Enable for complex tasks, disable for Flash-level speed
+```
+
+---
+
+## 2. GEMINI 3 FAMILY — Current Production Models
+
+### Gemini 3 Model Matrix (May 2026)
+
+| Model                  | Input Price/1M      | Output Price/1M         | Context | Status                         |
+| ---------------------- | ------------------- | ----------------------- | ------- | ------------------------------ |
+| Gemini 3.1 Pro         | $2.00 (>200K $4.00) | $12.00 (>200K $18.00)   | 1M      | GA                             |
+| Gemini 3.5 Flash       | $1.50               | $9.00                   | 1M      | GA (new in matrix Jun 2026)    |
+| Gemini 3.1 Flash-Lite  | $0.25               | $1.50                   | 1M      | GA (repriced from $0.10/$0.40) |
+| Gemini 3.1 Flash Image | $0.50               | $3.00 +$60/1M img out   | 1M      | GA                             |
+| Gemini 3 Pro Image     | $2.00               | $12.00 +$120/1M img out | 1M      | GA (new)                       |
+| Gemini 3 Flash         | $0.50 (audio $1)    | $3.00                   | 1M      | Preview                        |
+| Gemini 3.1 Flash       | $0.30               | $2.50                   | 1M      | Superseded by 3.5 Flash        |
+| Gemini 2.5 / 2.0 (all) | —                   | —                       | 1M      | Retired (after June 1)         |
+
+**CRITICAL:** Gemini 2.0 Flash, 2.0 Flash-Lite, and 2.5 Flash retired after
+June 1, 2026 — migrate any remaining calls to Gemini 3.5 Flash (standard) or
+3.1 Flash-Lite (budget). Note 3.5 Flash now costs $1.50/$9.00 — ~5x the old
+3.1 Flash; for cost-sensitive high volume, prefer 3.1 Flash-Lite ($0.25/$1.50)
+or Gemma 4 26B ($0.15/$0.60).
+
+### Gemini 3.1 Pro — What It Excels At
+
+```
+ARC-AGI-2 score: 77.1% (2x its predecessor)
+Best for: Complex reasoning, long documents, advanced code generation
+Pricing: $2.00/$12.00 per 1M tokens (>200K prompt: $4.00/$18.00)
+Context: 1M tokens (industry-leading for proprietary model)
+Use in ORIRO.ai: Strategy Builder, complex Campaign Exploder logic
+```
+
+### Gemini 3.1 Flash — The Sweet Spot
+
+```
+Best price:performance ratio for ORIRO.ai agents
+$0.30/$2.50 per 1M tokens (12-25x cheaper than Claude Sonnet)
+Speed: Faster than Pro, excellent quality
+Use in ORIRO.ai: All Flash Prompts, Brand Voice, Email Campaign
+Batch mode: 50% discount → $0.15/$1.25 for async workloads
+```
+
+### Gemini 3.1 Flash-Lite — Budget Champion
+
+```
+$0.10/$0.40 per 1M tokens — cheapest production-grade model
+GA since May 7, 2026
+Use in ORIRO.ai: High-volume content generation, embeddings
+No OpenAI equivalent at this price point
+```
+
+---
+
+## 3. SPECIALIZED MODELS — Image, Video, Audio
+
+### Imagen 4 (Current as of May 2026)
+
+```
+Models: imagen-4.0-generate-001, imagen-4.0-ultra-generate-001,
+        imagen-4.0-fast-generate-001
+Status: Preview (paid tier only)
+Pricing:
+  512px output: $0.045/image (747 tokens)
+  1K output:    $0.067/image (1120 tokens)
+  2K output:    $0.101/image (1680 tokens)
+  4K output:    $0.151/image (2520 tokens)
+Use in ORIRO.ai: Agent avatar generation, campaign image creation
+Notable: "significantly better text rendering" vs Imagen 3
+```
+
+### Veo 3.1 (Current as of May 2026)
+
+```
+Models: veo-3.1-generate-preview, veo-3.1-fast-generate-preview,
+        veo-3.1-lite-generate-preview
+Status: Preview (paid tier only for API)
+Pricing: 10 min of 1080p = ~$240
+Use in ORIRO.ai: Phase 4 video pipeline (Veo-powered Flash agents)
+D1 table: active model = veo-3.1 (confirmed)
+Note: Only charged if video successfully generates
+      Audio processing issues can prevent generation (no charge)
+```
+
+### Embedding Models
+
+```
+text-embedding-005: Current production embedding model
+  Use in ORIRO.ai: Agent knowledge base, semantic search
+  Free tier: available
+  Pricing: low (not published as of May 2026)
+
+multimodal-embedding-001: First multimodal embedding
+  Maps text, images, video, audio, PDFs → unified embedding space
+  New May 2026 — evaluate for future ORIRO.ai agent memory
+```
+
+---
+
+## 4. FREE TIER LIMITS (May 2026 — verified)
+
+**IMPORTANT:** Free tier limits changed significantly in 2026.
+Google cut quotas 50-80% in December 2025, more cuts April 1, 2026.
+
+```
+Google AI Studio (browser interface): Always free, no limits in UI
+API free tier (as of May 2026):
+  Flash models: higher RPM/TPM after May 7 2026 GA rollout
+  Pro models: 50 requests/day (cut from 100 on April 1)
+  All models: full 1M token context window
+  Data policy: FREE tier data may be used for model training
+               PAID tier: your data is NOT used for training
+
+Rate limit dimensions:
+  RPM = Requests Per Minute
+  TPM = Tokens Per Minute
+  RPD = Requests Per Day
+  Note: Check actual limits in AI Studio — numbers change monthly
+
+Free tier strategy for ORIRO.ai:
+  Development/testing: Free tier is sufficient
+  Production: Enable billing (spend caps prevent surprises)
+  Private user data: MUST use paid tier (data not used for training)
+```
+
+### Gemma 4 Free Access
+
+```
+OpenRouter free tier: Gemma 4 31B = $0/M tokens (rate limited)
+Google AI Studio:     Gemma 4 31B + 26B free for development
+Third-party hosted:   Replicate, Together AI, etc. — check current rates
+
+Self-hosting:
+  E4B: runs on modern laptop, phone, Raspberry Pi
+  26B MoE: RTX 4090 or Apple Silicon (24GB+ unified memory)
+  31B dense: H100 80GB (not for self-hosting on consumer hardware)
+```
+
+---
+
+## 5. VERTEX AI vs GOOGLE AI STUDIO
+
+```
+                    Google AI Studio        Vertex AI
+Development:        ✅ Instant access        Requires project setup
+Pricing:            Free tier + per-token   Per-token, enterprise SLAs
+Data privacy:       Free=training data      Private, GDPR compliant
+Production:         Small-medium workloads  Enterprise workloads
+Rate limits:        Lower (free tier)       Higher (paid)
+Non-global endpoints: N/A                  Pricing from July 1, 2026
+Use for ORIRO.ai:   MVP + development       Post-scale (phase 5+)
+
+ORIRO.ai decision: Google AI Studio API for now
+                   Vertex AI when user data privacy becomes critical
+                   or when rate limits become bottleneck
+```
+
+---
+
+## 6. ORIRO.ai — SPECIFIC MODEL DECISIONS
+
+### Current D1 google_ai_capabilities Table (Verified May 2026)
+
+```sql
+-- Active models in production:
+flash_primary:  gemini-2.5-flash    ← UPDATE to gemini-3.1-flash
+learn_primary:  gemini-2.5-flash    ← UPDATE to gemini-3.1-flash
+video:          veo-3.1             ← CORRECT ✅
+image:          gemini-3.1-flash-image  ← CORRECT ✅
+embedding:      text-embedding-005  ← CORRECT ✅
+
+-- Deprecated (update before June 1):
+gemini-2.0-Flash-Lite    → migrate to gemini-3.1-flash-lite
+gemini-2.5-flash         → migrate to gemini-3.1-flash
+```
+
+### Flash Prompts 1-8 — Right Model for Each
+
+```
+Flash Prompt 1 (Campaign Exploder):   gemini-3.1-flash ← $0.30/M
+Flash Prompt 2 (Brand Voice):         gemini-3.1-flash
+Flash Prompt 3 (Strategy Builder):    gemini-3.1-pro   ← complex
+Flash Prompt 4 (Content Calendar):    gemini-3.1-flash
+Flash Prompt 5 (Ad Copy):             gemini-3.1-flash
+Flash Prompt 6 (Email Campaign):      gemini-3.1-flash
+Flash Prompt 7 (Social Content):      gemini-3.1-flash-lite ← cheapest
+Flash Prompt 8 (Flash Dashboard):     gemini-3.1-flash
+```
+
+### Tier Model Integration
+
+```
+T1 (Free — Chrome Gemini Nano):
+  Uses on-device Chrome extension
+  Gemma 4 E2B equivalent performance
+  ORIRO pays $0 compute ✅
+  Limit: text only, no image generation
+
+T2 (Google Pro $0.99/mo):
+  Access to gemini-3.1-flash via user's Google API key
+  ORIRO pays $0 compute ✅
+  User brings their own key
+  Enable image generation via Imagen 4
+
+T3 (BYO Any Key $0.99/mo):
+  User supplies Gemini/OpenAI/Anthropic key
+  ORIRO passes key through — pays $0
+  Unlock Veo 3.1 for video agents
+
+ORIRO Cost Model: $0 compute on all tiers ✅ (confirmed correct)
+```
+
+### Gemma 4 Opportunity for ORIRO.ai
+
+```
+PROBLEM TO SOLVE: T1 users get limited Gemini Nano (Chrome only)
+OPPORTUNITY: Gemma 4 E4B runs on ANY device (no Chrome required)
+
+Option A: Offer Gemma 4 local mode for T1 users
+  - User downloads E4B (8B params) to their machine
+  - ORIRO agents run locally — no API calls, no cost
+  - 128K context, multimodal, function calling
+  - Privacy: all data stays on device
+
+Option B: Use Gemma 4 26B via OpenRouter free tier
+  - Free API access for development/low-volume
+  - $0.06/M for production
+  - Better than Nano for complex agent tasks
+
+Option C: Fine-tune Gemma 4 E4B on ORIRO agent personas
+  - Apache 2.0 = can fine-tune and redistribute
+  - One-time cost: $50-500 for LoRA fine-tune
+  - Inference cost after: same as base ($0)
+  - Create "ORIRO-Agent" specialized model
+
+RECOMMENDATION: Implement Option B for T1 power users
+Add Gemma 4 26B as T1 premium option alongside Gemini Nano
+```
+
+### Agent Memory Architecture (Gemma 4 Advantage)
+
+```
+Current ORIRO: gemini-2.5-flash for learn_primary
+Better with Gemma 4 31B reasoning:
+  256K context = full campaign history in one prompt
+  Native function calling = structured agent tool use
+  Thinking mode = step-by-step campaign strategy
+  Cost: $0.06/M vs $0.30/M for Flash (5x cheaper for reasoning)
+
+For Knowledge Base agents:
+  text-embedding-005 → embed user content
+  multimodal-embedding-001 → embed images+text together
+  Store in ORIRO's D1 or Cloudflare Vectorize
+```
+
+---
+
+## 7. MODEL SELECTION DECISION TREE
+
+```
+Is it a REASONING task (strategy, analysis, complex copy)?
+  → gemini-3.1-pro ($2.00/M) or Gemma 4 31B ($0.06/M)
+
+Is it a STANDARD agent task (content, email, social)?
+  → gemini-3.1-flash ($0.30/M)
+
+Is it HIGH-VOLUME, SIMPLE text generation?
+  → gemini-3.1-flash-lite ($0.10/M) or Gemma 4 26B ($0.06/M)
+
+Is it IMAGE GENERATION?
+  → imagen-4.0-fast ($0.045/img 512px) for speed
+  → imagen-4.0-ultra for quality
+
+Is it VIDEO GENERATION?
+  → veo-3.1 (paid tier, ~$240/10min 1080p)
+
+Is it EMBEDDING?
+  → text-embedding-005 (text only)
+  → multimodal-embedding-001 (text+image+video+audio)
+
+Is it ON-DEVICE (T1 Chrome user)?
+  → Gemini Nano (current)
+  → Gemma 4 E4B (future — better capability)
+
+Does user have their own API key (T2/T3)?
+  → Pass through their key, ORIRO pays $0
+```
+
+---
+
+## 8. CRITICAL MIGRATION — DO BEFORE JUNE 1, 2026
+
+```
+Deprecated models to replace:
+  gemini-2.0-Flash     → gemini-3.1-flash
+  gemini-2.0-Flash-Lite → gemini-3.1-flash-lite
+  gemini-2.5-flash     → gemini-3.1-flash (already mostly migrated)
+
+Files to update in ORIRO.ai:
+  D1 migration: infra/cloudflare/d1/migrations/
+  Update google_ai_capabilities table
+  Update any hardcoded model strings in apps/web
+
+Files to update in <project>-intel:
+  Update summarizer model string from old Gemini to gemini-3.1-flash
+  <project>-intel-src/main.py → grep for "gemini" model strings
+  <project>-intel-src/config.py → GEMINI_MODEL constant
+
+Files to update in the app router:
+  Update routes that reference old Gemini model strings
+```
+
+---
+
+## 9. BENCHMARKS — WHERE GEMMA 4 WINS
+
+```
+Gemma 4 31B vs competition:
+  MMLU Pro:        85.2% (better than most 70B models)
+  AIME 2026:       89.2% math reasoning
+  Codeforces ELO:  2150 (top-tier coding)
+  Arena AI rank:   #3 globally among open models
+  Arena AI score:  1452 (text only)
+
+Gemma 4 26B MoE (only 3.8B active):
+  Arena AI rank:   #6 globally
+  Arena AI score:  1441
+  Speed:           Comparable to 4B model
+
+This means ORIRO.ai agents can use Gemma 4 26B and get:
+  - Near-frontier reasoning quality
+  - 4B model speed
+  - 256K context window
+  - $0.06/M tokens (5x cheaper than Gemini 3.1 Flash)
+```
+
+---
+
+## 10. QUICK REFERENCE — MODEL STRINGS
+
+```python
+# Current production model strings (May 2026):
+GEMINI_PRO         = "gemini-3.1-pro"
+GEMINI_FLASH       = "gemini-3.1-flash"
+GEMINI_FLASH_LITE  = "gemini-3.1-flash-lite"
+GEMINI_FLASH_IMAGE = "gemini-3.1-flash-image"
+GEMINI_EMBED       = "text-embedding-005"
+IMAGEN             = "imagen-4.0-generate-001"
+IMAGEN_FAST        = "imagen-4.0-fast-generate-001"
+VEO                = "veo-3.1-generate-preview"
+GEMMA_4_31B        = "gemma-4-31b-it"         # OpenRouter free
+GEMMA_4_26B        = "gemma-4-27b-it"         # Google API
+
+# Old strings to REPLACE (deprecated June 1):
+OLD_FLASH_LITE = "gemini-2.0-flash-lite"    # ← REMOVE
+OLD_FLASH      = "gemini-2.0-flash"         # ← REMOVE
+OLD_25_FLASH   = "gemini-2.5-flash"         # ← UPDATE to 3.1-flash
+
+# <project>-intel summarizer update:
+APP_SUMMARIZER   = "gemini-3.1-flash"
+APP_SCORER       = "claude-haiku-4-5-20251001"  # unchanged
+```
+
+---
+
+## 11. GOOGLE I/O MAY 19-20 2026 — WHAT TO WATCH
+
+```
+Expected announcements (monitor and update ORIRO D1 + <project>):
+  - Gemini 3.2 Flash (rumored) → update flash_primary if released
+  - Gemini 3 Flash GA (been preview since Dec 2025)
+  - "Deep Think" mode rollout (advanced reasoning)
+  - Veo 4 (next video generation model)
+  - Imagen 4 updates
+  - Free tier rate limit changes
+
+Update procedure:
+  1. Check ai.google.dev/gemini-api/docs/pricing
+  2. Update D1 google_ai_capabilities table via migration
+  3. Update <project>-intel model strings
+  4. Update the app router routes
+  5. Test all Flash Prompts with new model strings
+```
+
+---
+
+_Read live from Google DeepMind, Google AI Developers, Vertex AI docs_
+_Prices and models verified May 19, 2026_
+_Next scheduled refresh: Monday May 25, 2026 (ClaudeSkills-GoogleAI-Refresh task)_
+
+---
+
+## 12. GOOGLE I/O 2026 — LIVE ANNOUNCEMENTS (May 19, 2026)
+
+**Source:** Tom's Guide live blog + BusinessToday + TechRadar — read May 19, 2026
+
+### NEW MODELS ANNOUNCED TODAY
+
+#### Gemini 3.5 Flash — JUST RELEASED
+
+```
+Status:    LIVE (announced keynote May 19, 2026)
+Described: "Strongest agentic and coding model"
+Benchmarks: Surpasses Gemini 3.1 Pro on:
+  - Coding
+  - Reasoning
+  - Multimodal understanding
+  - Terminal-Bench 2.1
+Use case:  Developers, businesses, complex AI tasks
+Speed:     Faster and cheaper than other advanced models
+Capabilities:
+  - Plan and organise tasks
+  - Build and improve solutions iteratively
+  - Handle practical business and software work
+  - Develop new apps
+  - Maintain/update codebases
+  - Prepare financial documents
+
+ORIRO.ai action: UPDATE flash_primary to gemini-3.5-flash
+<project> action:    UPDATE summarizer to gemini-3.5-flash
+Model string:    "gemini-3.5-flash" (confirm in AI Studio)
+```
+
+#### Gemini 3.5 Pro — COMING NEXT MONTH (June 2026)
+
+```
+Status:    Announced, not yet released
+Described: "Powerful version of Flash"
+ETA:       June 2026
+Capabilities: Across all Google products and platforms
+ORIRO action:  Plan migration from gemini-3.1-pro → gemini-3.5-pro
+               Update D1 when released
+```
+
+#### Gemini Spark — NEW AGENTIC AI
+
+```
+Status:    LIVE for AI Ultra users (US only initially)
+Tier:      AI Ultra ($100/month — new tier announced today)
+Built on:  Gemini 3.5 + Antigravity harness
+Type:      "24/7 personal AI agent"
+Connects:  Gmail, Docs, Slides, Google Workspace
+Functions:
+  - Declutter inboxes (analyze + summarize emails)
+  - Auto-archive newsletters
+  - Meeting summaries
+  - Structure info from Google apps
+  - Personal news briefings (follows stories over time)
+
+ORIRO.ai relevance:
+  - Gemini Spark = the agentic model ORIRO agents should aspire to
+  - Antigravity harness = Google's agent orchestration layer
+  - Study Antigravity 2.0 architecture for ORIRO agent design
+  - Flash Agents (ORIRO) + Gemini 3.5 Flash = competitive offering
+```
+
+#### Gemini Omni — VIDEO EDITING MODEL
+
+```
+Status:    Announced at I/O 2026
+Type:      Video generation + editing within Gemini app
+Powered by: Veo 3.1 (confirmed)
+Capability: Generate AND edit videos directly in Gemini chat
+ORIRO action: Monitor API availability for Veo phase (Phase 4)
+```
+
+#### Antigravity 2.0 — AGENTIC CODING SYSTEM
+
+```
+Status:    Showcased at I/O 2026
+Achievement: Created core components of an OS in 12 hours
+  - 93 separate sub-agents deployed
+  - 2.6 billion tokens generated
+  - Core OS framework in ~12 hours
+
+ORIRO relevance: This is the architecture for multi-agent coordination
+  Antigravity = orchestrator launching sub-agents
+  ORIRO's Flash Agents = similar sub-agent pattern
+  Study Antigravity for ORIRO's Phase 6 (1000+ agents)
+```
+
+### UPDATED MODEL PRIORITY FOR ORIRO.ai
+
+```
+IMMEDIATE UPDATES NEEDED (post I/O):
+
+1. D1 google_ai_capabilities — update today:
+   flash_primary:  gemini-3.1-flash → gemini-3.5-flash
+   learn_primary:  gemini-3.1-flash → gemini-3.5-flash
+
+2. <project>-intel:
+   APP_SUMMARIZER: gemini-3.1-flash → gemini-3.5-flash
+
+3. the app router routes:
+   Update any Gemini model string references
+
+JUNE 2026 (when Gemini 3.5 Pro releases):
+   flash_primary strategy: gemini-3.5-pro for complex agents
+   gemini-3.5-flash for standard agents (cost-optimized)
+
+NEW MODEL STRINGS (I/O 2026):
+   GEMINI_35_FLASH = "gemini-3.5-flash"    ← NEW ✅
+   GEMINI_35_PRO   = "gemini-3.5-pro"      ← COMING JUNE
+   GEMINI_SPARK    = AI Ultra tier only
+   GEMINI_OMNI     = Veo integration (monitor)
+```
+
+### SMART GLASSES — HARDWARE CONTEXT
+
+```
+Announced: Samsung 'Intelligent Eyewear' — fall 2026
+Also: Xreal Project Aura, other Android XR partners
+Powered by: Gemini (on-device via Gemini Nano equivalent)
+
+ORIRO opportunity:
+  Android XR glasses = new platform for ORIRO agents
+  Phase 6+: ORIRO agents accessible via smart glasses
+  Agent commands via voice → Gemini → ORIRO API
+```
+
+### I/O 2026 DEVELOPER KEYNOTE (May 20 — Tomorrow)
+
+```
+Session to watch: "What's new in the Gemma open model family"
+  - New Gemma additions will be revealed
+  - Practical deployment tools at scale
+  - Update this skill after May 20 developer keynote
+
+Expected from developer keynote:
+  - Gemma 4 updates or new Gemma 4 variants
+  - Vertex AI agent platform updates
+  - Google AI Studio new features
+  - Developer tools for Antigravity/agent building
+```
+
+### UPDATED QUICK REFERENCE (Post I/O 2026)
+
+```python
+# CURRENT production model strings (May 19, 2026 post-I/O):
+GEMINI_35_FLASH    = "gemini-3.5-flash"     # NEW — use this now
+GEMINI_31_PRO      = "gemini-3.1-pro"       # Until 3.5 Pro releases
+GEMINI_FLASH_LITE  = "gemini-3.1-flash-lite"
+GEMINI_FLASH_IMAGE = "gemini-3.1-flash-image"
+GEMINI_EMBED       = "text-embedding-005"
+IMAGEN             = "imagen-4.0-generate-001"
+VEO                = "veo-3.1-generate-preview"  # Omni coming
+GEMMA_4_31B        = "gemma-4-31b-it"
+GEMMA_4_26B        = "gemma-4-27b-it"
+
+# Deprecated (remove before June 1):
+OLD = ["gemini-2.0-flash", "gemini-2.0-flash-lite", "gemini-2.5-flash"]
+
+# ORIRO.ai D1 update SQL:
+# UPDATE google_ai_capabilities
+# SET model_id = 'gemini-3.5-flash'
+# WHERE capability_key IN ('flash_primary', 'learn_primary');
+```
+
+_I/O 2026 Day 1 keynote: May 19, 2026 — read live from Tom's Guide + BusinessToday_
+_Developer keynote Day 2: May 20, 2026 — UPDATE after Gemma session_