The Product Desk

Four independent sources this week converge on a single, urgent conclusion: single-provider AI lock-in is now both economically irrational and strategically fragile. The evidence is overwhelming.

The Numbers That End the Debate

Gemini 3.1 Pro Preview scored 57.2 on the Artificial Analysis Intelligence Index. GPT-5.4 Pro scored 57.0. The cost to run the same benchmark suite: $892 vs $2,950 — a 3.3x gap for effectively identical general intelligence. Open-weights GLM-5 hit 50 points at just $547. GPT-5.4 Pro earns its premium only in two narrow categories: coding (57 vs 56) and agentic tasks (69 vs 68 for Claude Opus 4.6). For everything else, you're paying 3x for equivalent output.

The era of single-provider AI is over. Raw intelligence is converging; cost-efficiency is diverging. Your competitive advantage comes from how you orchestrate models, not which one you use.

The Cautionary Tales

Meta invested $14.3 billion in Scale AI, poached CEO Alexandr Wang as Chief AI Officer, stood up a 100-person TBD Lab, and spent months building a flagship model code-named Avocado. The result? Avocado beat last year's Gemini 2.5 but couldn't match Gemini 3.0 on reasoning, coding, and writing. Meta's leadership reportedly discussed temporarily licensing Google's Gemini to power Meta's own AI products. A company that championed open-source AI with LLaMA is now considering renting a closed model from its biggest competitor.

Meanwhile, Ben Thompson documents Microsoft's three AI pivots in 18 months: from OpenAI exclusivity → infrastructure-around-models → bundling Anthropic's own integration into Copilot Cowork. The implication is stark: model makers beat wrappers at the integration layer, even when the wrapper has $200B+ in revenue and unlimited engineering resources.

The Winning Architecture

Practitioners have already converged on the answer. Power users report GPT 5.4 XHigh for production code, Opus 4.6 for design and planning, with tools like Droid and Pi supporting mid-conversation model switching. Adobe's Firefly integrated 25+ third-party models from Google, OpenAI, Runway, and Black Forest Labs — the Stripe play applied to creative AI. The common thread: the workflow layer wins, not the model layer.

The strategic question is no longer which model but which routing architecture. Map every AI-powered feature to a task category, assign a cost tier, and default to the cheapest model that meets your quality bar. GPT-5.4's cached token price of $0.25/M makes OpenAI quality accessible for template-heavy workflows — but only if you architect for cache hits.

The Ecosystem Is Forming Right Now

Four sources independently flagged the same phenomenon: a composable 'skills' ecosystem for AI agents is rapidly crystallizing, and the platform dynamics mirror Chrome Web Store, Shopify App Store, and Figma plugins from prior eras. The central marketplace: Vercel's Skills.sh, with daily updates and first-party skills from Anthropic (frontend-design), OpenAI, and Tailwind (ui.sh, imminent). One practitioner's visualization skill hit 200+ GitHub stars and 133K tweet views in a single day. Context Hub (chub), a CLI tool feeding API docs to coding agents, hit 5K GitHub stars and grew from ~100 to ~1,000 documentation files in its first week.

If your product has any AI-powered capabilities, you should be asking: should we package this as a skill and distribute through Skills.sh? The window for establishing platform position is measured in quarters, not years.

Developer Tools Are Being Redesigned for Agents as Primary Users

The shadcn/cli v4 release explicitly ships 'skills' to improve AI agent performance — not human developer performance. React's own documentation site now exports as Markdown (append .md to any URL) with a 'Copy Page' button, transparently designed for LLM consumption. MCP (Model Context Protocol) is crystallizing as the standard for agent-to-service communication — PropelAuth shipped an Integration MCP Server that lets a founder tell an agent 'set up auth with social login and match my brand colors' and get a working implementation. Claude Skills now packages 66 skills and 9 workflows around it.

Agent Infrastructure Is Becoming Enterprise-Grade

Three signals mark the shift from experimental to production:

1. Agent Identity: Teleport launched an Agentic Identity Framework with cryptographic identity, access controls, and observability — positioning agent security as a distinct category from traditional IAM.
2. Agent Economics: Ramp is building 'agent cards' — credit cards for AI agents to make autonomous purchases. Agents are becoming economic actors.
3. Agent Memory: Hindsight launched purpose-built agent memory. Superpowers launched coding agent workflow tooling. The stack is filling in.

The Defensibility Warning

One practitioner asked GPT 5.4 XHigh to reverse-engineer T3 Code's entire functionality — and it succeeded. T3 Code subsequently went open source, acknowledging that the interface layer has near-zero defensibility when an LLM can replicate it from a 30-minute demo. Defensibility has migrated up to data, workflow lock-in, skills ecosystem network effects, and enterprise trust. If your AI feature moat is 'better prompts and nicer UX,' you're building in the replicable layer.

An 8-level agentic engineering maturity model published this week frames each level as a 'huge leap in output' — and crucially, every model capability improvement amplifies gains at higher levels exponentially. The gap between your team and a competitor two levels above isn't closing with time. It's widening.

The Market Has Shifted Underneath You

The Sensor Tower State of Mobile 2026 report contains the single most consequential market data in today's briefing: 110 million Americans now use AI chatbots exclusively on mobile, up from 13 million at the start of 2024 — an 8.5x increase in 18 months. Users spent 48 billion hours in AI apps in 2025, nearly 10x the 2023 figure. AI app revenue tripled to $5 billion, and for the first time, non-game app revenue exceeded gaming — driven entirely by AI. OpenAI and DeepSeek captured ~50% of global AI downloads (up from 21% in 2023).

This isn't early adoption anymore. This is mass consumer behavior change. If your AI features live behind a desktop experience, you're building for the shrinking minority.

A critical nuance: Microsoft's data shows Copilot mobile users discussed health and fitness more than work and productivity. Mobile AI use cases are fundamentally different from desktop ones. Users aren't porting work habits to phones — they're inventing new AI habits in personal contexts. Meanwhile, all top 10 most-downloaded AI apps were general assistants. No vertical AI app has cracked mobile distribution at scale — both your biggest risk and your biggest opportunity.

Two Pricing Strategies Are Crystallizing — Pick One

In the same week, two competing pricing paradigms emerged:

The context behind OpenAI's move is defensive, not offensive: ChatGPT is reportedly seeing rising app uninstalls while Claude gains share. Bundling Sora is an attempt to re-sticky the platform. But video generation costs are dramatically higher than image generation, creating a structural margin problem. Either OpenAI raises prices (more churn), implements aggressive caps (undermines the value prop), or burns capital. Watch Q2 pricing moves like a hawk — whatever they choose becomes the template.

Adobe set the contrarian standard: unlimited AI generations for paid Photoshop users and Firefly subscribers. Natural language is now the primary editing interface — removing objects, changing lighting, adjusting backgrounds via text prompts. AI Markup adds sketch-to-edit. Users are being trained right now to expect this as baseline UX.

Google Maps Collapses the Local Discovery Funnel

Google's 'Ask Maps' enables natural language queries like 'cafés with short lines where I can charge my phone' — personalized by history, powered by 500M+ contributors. This collapses discover → research → decide into a single AI-mediated interaction. Any product relying on fragmented local discovery (Yelp → decide → go) just lost a step in the funnel. More Gemini integrations expected before Google's May developer conference.