Last quarter, a mid-sized team in Barcelona shipped an on-chain research agent. The first draft of its weekly market report cost them forty cents in inference. The final, shipped version β after the agent checked its own citations, retried twice, and reconciled three conflicting data feeds β cost eleven dollars and sixty cents. The model had gotten cheaper. The invoice had gotten larger.
That gap is the story the market refuses to price. We spent two years celebrating the collapse of per-token costs; we spent almost none of that time asking what an agent does with tokens after the first answer. The bottleneck in agentic AI is not generation β it is the refinement loop, the recursive self-verification that quietly dominates the bill. For anyone trying to settle agent work on a public ledger, that loop is not a footnote. It is the entire accounting problem, and it is where the next narrative collapse is being underwritten.
To hunt the truth, one must first bury the hype. The hype was cheap tokens. The truth is expensive verification.
For a blockchain audience, this should feel like dΓ©jΓ vu, because we have lived this arc before. The 2017 ICO cycle priced utility as though the token were the product; the 2020 DeFi summer priced liquidity as though the pool were the business. Each narrative collapsed not because the technology failed, but because the metering β the mechanism that assigned value to work β was fiction. When you cannot measure what you are selling, you are not selling it; you are borrowing belief against it.
AI agents are walking into the same room. The industry is migrating from charging by token toward charging by compute, by workload, by outcome. Pricing research from Orb suggests the overwhelming majority of AI vendors now run hybrid models, with usage metering near-universal and pure outcome pricing a rounding error β roughly 3.8%, and shrinking. Stripe's data points to the same restlessness: most usage-billed AI firms have already repriced at least once. The direction is real. And in a bear market, this matters less as an opportunity than as a survival question β which protocols are bleeding, and whether the assets parked inside them are safe.
The numbers circulating alongside the trend, however, deserve suspicion. A Haiku 5.5 at ten cents per million input tokens; a Mistral ML4 that exists in no naming lineage I have ever audited; reports dated to months that have not happened. I flag these not to be pedantic, but because a pricing narrative built on unverifiable figures is itself a metering failure β the exact information asymmetry the analysis claims to expose, turned back on its author. To hunt the truth, one must first bury the hype, and the hype here wears a spreadsheet.
What survives the noise is directional, and I think correct: the unit of account is shifting from tokens toward something closer to work.
Here is where the crypto angle stops being a metaphor. If agents are to transact on-chain β paying for compute, settling for results, accruing reputation β then the meter becomes a trust primitive, and trust primitives are precisely what blockchains claim to provide. But the refinement loop breaks the naive model.
An on-chain meter must answer three questions the token model never asked. First, how do you normalize compute across heterogeneous hardware β an H100, a TPU, a domestic accelerator β when their units are not commensurable? A step on one chip is not a step on another, and a ledger cannot settle a fraction it cannot define. Second, how do you attribute cost when an agent's answer depends on a KV-cache shared across a hundred prior calls, where the marginal token is nearly free and the amortized token is not? Third, and hardest: who bears the cost of a wrong answer that consumed real compute?
Based on my audit experience across DeFi protocols, I have watched this pattern before. Every system that tried to price work without pricing verification eventually priced verification implicitly β through slashing, through disputes, through opacity. The lesson from liquidity provision is that incentives do not vanish when you refuse to meter them; they migrate into the spread. Agent billing is no different. If you cannot see the refinement loop, you are paying for it somewhere you cannot see.
There is an older precedent, and it is uncomfortable. Bitcoin's proof-of-work is the original compute meter β the first system to price raw work. And its lesson is that metering compute concentrates. After the fourth halving, miner revenue collapsed and hash power drifted toward a handful of pools; the difficulty adjustment stayed decentralized in theory while the work concentrated in practice. Agent metering is walking the same road. The measure looks neutral; the ability to control the measure is not.
The deeper technical truth is unflattering for the put-everything-on-chain crowd. The refinement loop exists because single-pass reliability is insufficient. You do not fix that with a better billing unit; you fix it with deterministic verifiers β unit tests, type checkers, compiler feedback β that replace the model's self-doubt with something a ledger can adjudicate. That is a capability problem wearing a pricing costume. And notice what the billing debate never mentions: caching, prefix reuse, batching β the real technical levers that compress refinement cost. They are omitted because they are hard, and a new meter is easy to sell. The enterprise symptom is already visible: most organizations report blowing past their AI budgets, and the refinement loop is a leading cause β a signal that governance, not generosity, is the missing layer.
Now the contrarian turn, and it cuts against my own industry's enthusiasm. The prevailing crypto narrative insists outcome-based agent pricing is the future β pay only for solved tickets, resolved queries, delivered code. It sounds fair. It is also, I suspect, the least likely model to dominate, and the reason is behavioral, not technical.
Outcome pricing forces the supplier to absorb execution risk. If the agent burns compute and fails, who eats it? The vendor. That compresses margin and, worse, makes revenue unpredictable β and unpredictable revenue is poison to the valuation story that funds the entire stack. So suppliers carry a structural incentive to keep the meter opaque and the unit smooth. Metering opacity is not a bug in the pricing model; it is the moat. The harder a bill is to audit, the more room it holds.
This mirrors the Data Availability debate I have circled for years. Everyone built dedicated DA layers for rollups that never generated enough data to need them. Everyone is now building outcome-pricing rails for enterprises that will quietly prefer the predictability of seats and tokens. Infrastructure arrives before demand; demand arrives wearing a different label. The real migration is not toward outcome pricing β it is toward metered opacity: usage-based bills that feel transparent and settle somewhere the buyer cannot fully verify. Trust is the new collateral, and it is scarce precisely where it is most needed.
So watch the meter, not the model. The protocols that win agent settlement will not be the ones with the cheapest tokens; they will be the ones that make verification cheap, portable, and auditable across hardware that refuses to agree on units. The refinement loop is where the value hides β and where the next collapse is being quietly financed. Ask of every agent invoice the question no vendor wants to answer: who is paying for the second guess?