The Cost Per Task: What the Ledger Reveals When Token Prices Lie

Over the past seven days, a mid-sized decentralized inference marketplace settled 4.1 million tasks and emitted rewards for 61 billion tokens. The verified completion rate was 38 percent. Nobody in the protocol's governance forum mentioned the other 62 percent, because failure, on-chain, still pays.

I found this by accident. I was pulling settlement logs for an unrelated piece of research on verifiable compute, and the numbers would not reconcile. The marketplace's dashboards measured input and output tokens with admirable precision. They measured nothing else. Not whether the task succeeded. Not whether a human had to redo it. Not whether the buyer received anything usable. The protocol had built an impeccable accounting system for a metric that no longer reflects what anyone actually buys.

This is not a story about one project's design flaws. It is a story about a pricing assumption the entire industry — crypto and conventional AI alike — has been quietly inheriting from two years of price war. And it is a story about what happens when that assumption meets a market where the settlement layer is public, permanent, and unforgiving.

Here is the claim that started this line of thinking. An unnamed analysis, circulating in industry channels last month, argues that United States AI models are often cheaper per task than their Chinese rivals. The assertion contradicts the dominant narrative of the past eighteen months: that DeepSeek, Qwen, GLM, and Doubao have collapsed the price of inference and left Western labs selling a commodity at five to twenty times the going rate.

On the surface, the price data supports the popular story. DeepSeek-V3 has been quoted around $0.27 per million input tokens and $1.10 per million output tokens, with further discounts on cache hits. GPT-4o sits nearer $2.50 and $10. Claude Sonnet, $3 and $15. Multiply tokens by unit price and the Chinese models win by an order of magnitude. That arithmetic is not in dispute.

What is in dispute is whether tokens are the right unit. The analysis — and I want to be precise about how thin the evidence is, because there is no named author, no model list, no task set, no methodology, and no sample — proposes that enterprise buyers are shifting their pricing anchor from cost per token to cost per task, or cost per business outcome. On that measure, the claim goes, the ranking can invert: a model with a higher unit price but a higher first-pass success rate generates fewer retries, fewer escalations, and less human cleanup, and therefore costs less to actually use.

I cannot verify that claim. Neither can anyone reading the same brief. That unverifiability is itself the most interesting thing about it — because the same gap exists inside the crypto inference markets I spend my working life auditing, where the numbers are public and the reconciliation still does not happen.

The technical core of this argument is a single equation, and it is worth writing out in full because almost nobody in either industry actually computes it:

cost per task = (input tokens × input price + output tokens × output price) × (attempts ÷ success rate) + human review cost

Three variables live inside that expression, and only one of them appears on a pricing page.

The first is token efficiency: how many tokens a model consumes to complete a given task. The second is reasoning overhead — the chain-of-thought tokens that reasoning models emit before answering. The third is first-pass reliability: the probability that the model completes the task correctly without retry, human correction, or escalation. Unit price is the only input any vendor publishes. The other three determine whether the published price means anything at all.

Consider what happens when a reasoning model meets a trivial task. DeepSeek-R1 class models are cheap per token and generous with their thinking. On a classification or extraction job that a smaller model would resolve in 200 output tokens, a reasoning model may emit several thousand tokens of deliberation. Its unit price advantage survives the multiplication. Its cost-per-task advantage does not. The same mechanism works in the opposite direction for frontier American models on hard tasks. If o-series or Claude extended-thinking models resolve a complex coding or multi-step agent task on the first attempt where a cheaper model requires three or four passes, the expensive model is the frugal one.

The insight is not that Chinese models are expensive. It is that token price is a denominator with no numerator attached, and enterprise procurement has started to notice.

I want to be honest about the limits of this reasoning. The task-type dependency is decisive, and the brief that prompted this essay buries it in a single adverb. "Often cheaper" is a conditional claim wearing the costume of a general one. On classification, extraction, summarization, and high-volume content work — the tasks that constitute the majority of enterprise inference volume by count — a model priced at one-tenth the unit cost will almost certainly win on cost per task. The inversion only becomes plausible on complex reasoning, code generation, and multi-step agentic workflows. Any analysis that reports a single aggregate number without stratifying by task class has told you more about its task selection than about the models.

Then there is the variable that both the unnamed analysis and most crypto inference protocols omit entirely: human review cost. Every deployed AI system I have audited has a human in the loop somewhere — a reviewer, a fallback operator, an engineer triaging failures. That labor is the single largest line item in most real deployments, and it is the one that never appears on a token dashboard. When I spent two hundred hours mapping voting centralization risks in Compound's governance mechanism back in 2020, the most useful finding was not about the code. It was that the social contract around the code — who reviews, who escalates, who bears the cost of a bad outcome — determined whether the mechanism actually worked. Smart contracts do not eliminate human error; they relocate it, and they relocate its cost with it.

We audit the logic, for humans will always err — and the cost of that erring is the part of the equation that no one publishes.

This is where the story sharpens, because a public settlement ledger makes the omission legible. Decentralized inference marketplaces — the DePIN networks selling GPU time against on-chain payments — have inherited the per-token pricing model wholesale. It is an understandable inheritance. Tokens are metered, verifiable, and cheap to attest. Success is subjective, expensive to verify, and contested. So the protocols price what they can measure. Miners are rewarded per token served. Buyers are charged per token consumed. The incentive gradient points, with mathematical inevitability, toward token production rather than task completion.

Code is the only law that does not sleep, and it enforces whatever you actually wrote.

I have watched this pattern before, in a different market. When I reviewed the digital collectibles sector for my essay "Pixels Without Principles," the structural problem was identical in shape: a market with primary issuance and no functioning secondary layer, where the thing being sold had no mechanism to prove it was worth holding. The tokens sold. The utility did not accrue. What looked like a market was a sequence of one-off transactions dressed in market clothing. The DePIN inference sector is running the same experiment with compute. Tokens are served, emissions are distributed, and the question of whether any buyer received a usable result is left to a dashboard that does not ask.

I will go further, because I have spent enough time inside protocol code to say it plainly. A meaningful share of the networks marketing themselves as decentralized inference are cloud resellers with a token bolted on — the compute is centralized, the verification is nominal, and the decentralization lives in the governance token rather than the stack. This is the same rebranding dynamic I have documented elsewhere in this industry: projects migrating to whatever narrative is currently funded, keeping the architecture and rewriting the pitch deck. The pattern is not new. It is simply arriving now in the one sector where the underlying economics are genuinely hard, and where a lazy pricing model is therefore genuinely dangerous.

I seek the signal amidst the noise of the crowd, and the signal here is that per-token settlement measures activity, not value.

There is a technical reason the protocols price tokens, and it deserves a fair hearing: token counts are cheap to verify and outcomes are not. Confirming that a model produced a correct answer requires either a trusted evaluator, a deterministic test harness, or a cryptographic proof of computation — and the third option, the one that would make outcome pricing trustless, is still slow and expensive relative to the task it certifies. So the industry defaults to the measurable proxy. I understand the engineering constraint. I do not accept it as permanent, and I have watched the cost of verification fall year over year while the cost of not verifying compounds in silence.

The open-source question cuts across all of it, and the brief that started this essay does not mention it once. DeepSeek and Qwen ship weights. When a model's weights are open, the token price vanishes as a concept — there is no per-token invoice, only the depreciation of the hardware you run it on and the electricity you feed it. The cost-per-task calculation does not disappear; it changes its inputs entirely, from a vendor's price list to your own utilization curve. Any comparison of American versus Chinese model cost that prices only hosted APIs has excluded the deployment mode in which a large fraction of technically capable buyers actually operate. Open source is a covenant, not just a license — and a covenant changes who bears the risk.

Which brings me to the compliance layer, and to the reason I believe the headline comparison is largely theater. For most United States enterprises, Chinese frontier models are not a purchasable option at all. Data sovereignty rules, procurement policy, export controls, and internal risk review remove them from the menu before cost is discussed. The real comparison inside an American procurement cycle is not DeepSeek versus GPT; it is one American vendor versus another. The cross-border cost comparison, however carefully computed, describes a choice most buyers are not permitted to make.

And the compliance apparatus itself deserves scrutiny. In my experience auditing KYC and procurement flows, the burden falls almost entirely on legitimate participants. Determined actors route around verification with trivial effort; honest users submit documents, wait for review, and absorb the friction. The compliance cost is real, it is substantial, and it is paid by exactly the people the system was not designed to stop. Any total-cost-of-ownership model that omits this line is not a cost model. It is a price list.

Here is the pragmatic test I apply to claims like this one, and it fails in an instructive way. If cost per task were genuinely becoming the enterprise standard, we would see it in procurement documents. We would see RFPs specifying cost per resolved ticket, per merged pull request, per completed workflow. We would see vendors publishing outcome-denominated price sheets. We would see, in the on-chain inference markets, protocols competing on verified task completion rather than emitted tokens. I have looked for all three. I have found fragments — a handful of vendors experimenting with per-seat and per-outcome billing, one or two marketplaces adding optional verification — and nothing that resembles an anchor shift.

The blind spot is not that the analysis is wrong. It is that the analysis is unverifiable by design, and that its most likely origin is a party with a direct interest in the conclusion. An unnamed study finding that American models are cheaper is a useful artifact for American labs facing a price-shock narrative. That does not make it false. It makes it unauditable — and in a market whose entire value proposition is verifiability, unauditable claims should be treated as marketing until proven otherwise.

Faith in people is costly; faith in math is free. The trouble is that nobody has written the math for cost per task yet.

The second blind spot is symmetry. Crypto has spent three years building public settlement layers — the exact infrastructure that would make cost-per-task measurement possible — and has used them to meter tokens instead. We have the ledger. We are not keeping the books.

The measurement that matters is not how many tokens a system consumed. It is how many problems it actually closed, and at what human cost. Somewhere in the next cycle, a protocol will publish verified task-completion rates on-chain, priced per outcome, and the rest of the market will have to answer for what its dashboards were hiding. Hype burns out; robustness remains in the ledger. The only question is who writes the first honest entry.