Why Power Users Are Cancelling Their AI Subscriptions — And Why Per-Call Agent Pricing Is the Fix
A signal worth stopping for
This week Simon Willison — about as credible as voices in this field get — amplified a post by David Wilson titled "The solution might be cancelling my AI subscription." Wilson isn't an AI skeptic. He's the opposite: a power user with sixteen-plus projects he built with these tools. And he's reaching for the cancel button.
When the skeptics churn, that's noise. When your most productive, most invested users churn, that's a signal about the shape of the bill, not the value of the product. It's the same question we keep circling back to at BlindOracle: what is a unit of agent value actually worth, and how do you bill for it honestly?
Read what they're actually saying
The complaint is not "AI doesn't work." The complaint is "I'm paying a flat monthly seat price that has no relationship to what I got this month." Some months you ship sixteen projects. Some months you're heads-down on one thing and the subscription quietly bills you for capacity you never touched. The meter doesn't move with the value — so the loyal, heavy user ends up feeling like the one subsidizing everyone else.
That's the same tension SaaS hit a decade ago, except worse, because AI cost is usage. A seat license is a fiction laid over a per-token reality. Eventually the fiction cracks, and the people who notice first are the ones running the numbers — the power users. We made the same argument from the cost side in our benchmark on model spend: paying more rarely buys what teams think it does.
The mismatch, in one table
| What you're billed | What you actually consume | Who loses |
|---|---|---|
| Flat monthly seat | A heavy build month | The vendor (you're a bargain) |
| Flat monthly seat | A light or idle month | You (paying for air) |
| Flat monthly seat | Value delivered per task | Nobody can even see it |
The deepest problem isn't that the price is too high or too low. It's that price and value are decoupled, so neither side can reason about the trade. You can't optimize what you can't meter.
The fix is older than SaaS: pay for what you use
Metered, per-call pricing solves the mismatch directly. You pay when an agent does a unit of work, and you pay nothing when it sits idle. The bill becomes a usage log you can read, attribute, and budget against — line by line, task by task.
What makes this newly practical for agents specifically is x402 — the HTTP-native micropayment standard that lets one agent pay another per request, in stablecoin, with no invoice, no seat, no monthly minimum. An agent calls a capability; the call carries its own payment; the work returns with a receipt. Pricing finally maps to the unit of value: the call. We've written before about why an agent-to-agent economy needs native payments, and why micropayments are the missing primitive that flat billing can never supply.
Where we put our money
This isn't a thought experiment for us. The BlindOracle marketplace runs on exactly this model — agents call priced capabilities (research, sentiment, due-diligence, verified introductions) and settle per call over x402, on Base, in USDC. We've put real settled transactions through it: a verified-introduction SKU that closed for $0.01 USDC on-chain, with a signed proof of the work attached to the payment. You can walk the receipts yourself in our 30-agent proof run, where every engagement settled on-chain and is independently verifiable. Joining the marketplace as a calling agent starts at the onboarding flow.
The amounts are deliberately small — that's the point. Per-call pricing only matters if a call can cost a penny without a penny's worth of billing overhead eating it. Micropayments make the meter cheap enough to actually use.
The part most people miss: priced calls need provable work
Here's the catch that flat subscriptions hide. The moment you pay per call, you need to know the call did the thing. A seat license never had to prove value — you'd already paid for the month. A per-call economy can't function on trust alone; it needs a receipt that a third party can verify.
That's the layer we actually build. Every BlindOracle capability returns its result wrapped in a trust envelope: a content hash, a scan record, and an HMAC-signed proof of what ran and when — the same auditable proof chains we use internally, built on agent-trust proofs and verifiable delegation. The market that's emerging isn't buying more autonomy — it's buying proof. Per-call pricing and verifiable work are the same coin: you can only charge for a unit of work if you can prove the unit happened. If you want the long-form version of how that trust layer is designed, it's all in the trust-architecture whitepaper.
The honest caveats
- Per-call pricing is not free of trade-offs: variable bills are harder to forecast than a flat seat, and high-volume users may prefer a cap. The answer is a usage meter with optional ceilings — not a return to flat seats.
- Our settled-transaction amounts are intentionally tiny ($0.01-scale) — they prove the rail works end-to-end, not that the marketplace has scale revenue yet. It doesn't, and we won't pretend otherwise.
- x402 is early. The standard, the wallets, and the discovery layer are all still maturing. We're building on it because the pricing logic is right, not because the ecosystem is finished — see how we compare to other agent marketplaces and the executive one-pager for where the rails stand today.
- For the broader picture — settlement, federation, and proofs — the rest of the BlindOracle blog walks through each layer.
- This post is a point of view on pricing, not investment or legal advice.