Utilize all AI subscriptions while you can

AI is becoming electricity.

Five years ago “do you use AI?” was a meaningful question; today it’s like asking whether someone uses Google search. The holdouts are a real cohort, but a shrinking one. The rest of us are quietly absorbing AI into email, code, research, scheduling, decisions. The shift from “tool you reach for” to “substrate you stand on” is most of the way done.

When something becomes substrate, the conversation moves from features to pricing. And the pricing on AI right now is a promotional period dressed as a business model.

The shape of the customer base

Think of it like PG&E. The customer base splits into two extremes and a soft middle:

Cohort	How they use AI	What that means
Baseline	Free tier; occasional question. The habit never formed — the way some people never really took to Google Search either.	Permanent cohort, larger than the industry pretends.
Power user	Maxes the window inside an hour, then sits and waits for the reset. Bottleneck is time, not money.	Already paying per token via the API (by maxing out subscriptions). Subscription is a torture, not a deal.
Middle	Uses AI more than occasionally, less than constantly. Free tier doesn’t cut it; Subscription is “technically fine.”	What subscription pricing is designed to extract from — and the least stable cohort.

The middle is the unstable cohort. It leaves in both directions:

Some run the math on the API and notice the per-token price is the same or cheaper, with no quota window to live around.
Others run a different math and notice AI doesn’t actually move their work or life enough to justify the bill, and drop to free.

Either way, the subscription tier hollows out from the middle.

Why subscriptions usually work, and why this one doesn’t

A subscription as a business model is durable when three conditions hold:

Most people need the thing. Wide demand keeps the funnel full.
Casuals subsidize heavy users. The light cohort pays the same flat rate but consumes a fraction of the cost, propping up the unit economics.
No clean pay-as-you-go alternative. Netflix doesn’t sell single episodes. Gyms don’t price single workouts at break-even. The subscription has to be the only practical door.

AI fails the third condition. The API door is right there, and for the heavy user the subscription isn’t a deal — it’s a constraint dressed as one. Three things converge:

The window caps your output. You stop shipping when the quota runs out, not when the work is done.
The reset clock owns your schedule. You rush to drain the quota before reset so you don’t “waste” what you paid for, then race to start again the second the new window opens. It’s a constant low-grade mental tax that bleeds into real work-and-life arrangement.
The provider can’t actually give you more than you pay. At best the subscription nets out to compute equal to its price — any more would be a subsidy the company can’t sustain. So the upper bound on a flat-rate plan is the same as buying the API credit yourself, minus the stress.

So today’s subscription tier is propped up mostly by users who haven’t run the math, plus marketing inertia.

The bundle isn’t the moat

The natural counter is that the subscription isn’t selling tokens — it’s selling a product: memory, file upload, web search, image generation, mobile app, integrations with Gmail and Drive. Switch to raw API and you lose most of what you actually use. That defense is real. It’s also shrinking faster than the providers would like, for three reasons.

1. The bundle is mechanically just wrappers around the same API.

Each “feature” is a thin layer on top of the model:

Bundle feature	What it actually is
Memory	A database lookup, prepended to the next prompt
File upload	Parse, chunk, embed, retrieve
Web search	A search API the model calls as a tool
Image generation	Another model behind another endpoint
Mobile app	A thin client over the same chat API
Gmail / Drive	OAuth + REST calls, like any third-party integration

None of this is hard. The model is the expensive part of the stack; the wrappers are the cheap part.

2. The wrappers are getting trivially cheap to build.

As model capability rises, the cost of building these wrappers collapses — to a weekend project for a developer, or a cheap third-party tool for everyone else. Open agent frameworks like OpenClaw wire up the full set (file, web, calendar, mail) against any API endpoint, in less time than it takes to read the docs. This is part of why incumbent SaaS valuations are taking a beating right now: the bundle that used to be the moat is now the cheapest layer in the stack to replicate.

3. The model — the actually expensive part — is available through every door.

You can reach the same Claude or GPT through the official chat, the API, or any of the dozens of third-party wrappers that hit the same endpoint. The provider doesn’t get to gate the model behind the bundle, only price the two doors differently.

Stack those three and the implication is sharp: first-party bundles have to be seamlessly excellent to keep paying customers, not merely good enough. Any rough edge — slow voice mode, mediocre file handling, an awkward mobile flow — and a third-party alternative wins on that specific feature, even if no single competitor matches the breadth of the bundle. Death by a thousand thin cuts.

The window is closing

Stack the pieces and the picture is uncomfortable:

Power users cost more than they pay. The flat rate caps revenue; the heavy compute doesn’t cap.
The middle leaves in both directions. Some defect up to the API, others defect down to free.
The bundle is replicable. Wrappers around the same model — and cheap to build.
Capital is funding the gap. Providers are buying market share with investor money, the same playbook as ride-share and food delivery a decade ago.

That’s a promotional period, not an equilibrium. The fix looks familiar: smaller windows, slower models, higher tiers, or a quiet swap to a cheaper variant on the subscription. It’s already starting — Anthropic just cut off OpenClaw and similar wrappers that turned one Pro plan into a generic agent backend.¹ That was the most visible move; the quieter ones — rate limits, model swaps — come next.

Three tiers, not one

Here’s my bet. When the flat-rate tier breaks up, it splits into three, sorted by how much each user is worth to the provider:

Free / baseline. Cheap models, no fee. The exchange is data: light users get the substrate at $0, providers get conversation logs to keep training on. The cohort that never converts gets to stay — just on a model that’s a generation behind.
Subscription. A flat entrance fee buys access to the frontier model and the bundled features (memory, file upload, search, mobile app), but not unlimited compute. Past a budget, you pay per token — the way phones meter overage. The “all you can eat” framing dies; subscriptions become hybrid plans.
Power users. Full migration to API keys plus a custom workflow they build once and run forever. They were never the subscription’s customer — they were its cost center. They walk away cleanly the moment the math crosses, and never come back.

None of this is novel — every metered-vs-flat-rate market has eventually landed somewhere like this. Cell phones did it. AWS did it. Streaming is doing it now. I’d bet AI converges on the same shape, just faster.

So the title isn’t a marketing flourish. If you’re paying for an AI subscription today, the rate plan you’re on is a transfer from investors and from people who haven’t run the numbers yet. Neither source is permanent. Use the model. Build the workflows. Move the work that’s worth moving.

Extract the subsidy while it’s still there!

Anthropic cut off subscription-based OAuth tokens to OpenClaw and similar wrappers on April 4, 2026, citing economic sustainability — a Claude Max plan costs $200/month while a 24/7 OpenClaw instance can consume $1,000–$5,000 in actual compute. See VentureBeat and The Register. ↩