Contrary to the narrative, a 90% price cut is not a 90% cost reduction.
On the most recent small-model pricing update from a major AI lab, the input line moved from $1.00 per million tokens to $0.10. The output line moved from $5.00 per million tokens to $0.50. Both fell by exactly one order of magnitude. The framing was uniform across every aggregator that carried the sheet: the cheapest capable model yet.
Then came the second figure. Average customer cost, the vendor stated, fell by approximately 75%.
Two percentages. Ninety and seventy-five. On their face they do not reconcile, and they should โ they describe the same event. Divide the cost ratio by the price ratio: 0.25 divided by 0.10. The quotient is 2.5. The average task now consumes roughly 2.5 times more tokens than it did on the prior version. The vendor's gloss โ "slightly higher token counts" โ is off by a factor of 2.5. That is not a rounding note. It is the entire economic story, inverted. And it landed in the crypto industry's lap without anyone in crypto noticing.
Why this matters to anyone holding an on-chain position is not obvious from a press release. It becomes obvious the moment you look at what is actually being built.
The past eighteen months have produced a category of protocols I will call inference-dependent: autonomous wallets, agent-managed vaults, on-chain trading strategies, and the broader "AI agent" frameworks that route capital based on model output. These systems do not run a model. They call one. Every decision โ rebalance, exit, hedge, liquidate โ is an API request with a token bill attached. The protocol's unit economics are therefore a function of exactly two variables: the model's price per token, and the model's token consumption per decision.
The industry has spent a year optimizing the first variable and ignoring the second. Marketing pages quote price-per-million-token as if it were a fixed cost of production. It is not. It is one factor in a product, and the other factor just moved by 150%.
That is the context in which a small-model pricing sheet becomes a crypto story. When I audited the first major autonomous-wallet protocol last year, I found a logical race condition in the reinforcement-learning reward function that permitted infinite minting under specific market conditions. The fix was deterministic. The interesting part was the cost model underneath it. The protocol's backtest assumed a flat inference cost per decision. The mainnet assumption broke within two weeks of launch โ not because the price changed, but because the consumption did.
Trust is a variable; proof is a constant. The price is the variable everyone watches. The consumption is the constant nobody measures.
The cycle itself is familiar. A capability demo โ usually a benchmark chart โ is repackaged as a product, the product is repackaged as a token, and the token is priced against a total addressable market that assumes the underlying compute is nearly free. When the compute is not free, and when its cost is misreported, the token price is anchored to a number that does not exist. This is not a new failure mode. It is the same failure mode as unbacked yield, wearing a different interface.
The effort setting is the real disclosure
The sheet did contain one genuine technical disclosure, buried under the price table: an adjustable "effort" setting. It is the most substantive item in the release, and it went almost unremarked.
An adjustable effort control is test-time compute scaling, exposed to the user. It means the model is not a pure next-token predictor; it is a hybrid that allocates variable reasoning budget per request. The implication is mechanical. The same prompt can cost five times more or five times less depending on where the dial sits. When a protocol quotes "cost per decision," it is quoting a number that the model configuration can move by an order of magnitude without any change to the underlying price sheet.
For an on-chain agent, this is not a feature. It is an unbounded liability. An autonomous vault that rebalances on a schedule cannot have a cost function that floats with a hidden effort parameter. If the effort setting is not pinned, the protocol's cost per decision is not pinned. A cost model with an unpinned variable is not a cost model; it is a range. And a range is not something you can underwrite a token against.
The 2.5x and what it does to a backtest
Return to the arithmetic, because it is the load-bearing wall of this entire analysis.
Price ratio: 0.10. Cost ratio: 0.25. Token consumption ratio: 2.5.
This is the cleanest forensic artifact in the release, and it was disclosed by accident. A vendor advertising a 90% cut has no incentive to publish the figure that halves the benefit. The 75% number leaked because it is the number customers actually feel, and it contradicts the marketing number.
Propagate it. Take a hypothetical agent-managed vault whose strategy fires 1,000 model decisions per day. Under the old model, at, say, 2,000 tokens per decision, the daily bill is bounded and predictable. Under the new model, the price per token falls 90% โ but the token count per decision rises to 5,000. The net saving is not 90%. It is the price ratio times the consumption ratio: 0.10 times 2.5, or 0.25. A 75% saving. Exactly the figure the vendor reported, which is precisely why the figure is credible.
A 75% saving is still a saving. But a protocol that underwrote its token model on a 90% reduction has now mispriced its operating cost by a factor of 1.6 against its own assumption. Over a year, on a vault with meaningful notional, that is not a rounding error. It is the difference between a strategy that nets positive and one that bleeds โ and the bleed is invisible on the dashboard until the vault is underwater.
The benchmark jump is real, and that is the problem
The release also reported a capability step-change on agentic benchmarks. Terminal-Bench 4 moved from 0% on the prior small model to 39.2% on the new one. OSWorld 2.1 was cited at 48.9%.
A single-generation move from 0% to 39.2% is not capability growth. Capability grows in curves, not cliffs. Two explanations are consistent with the data. The first is a post-training paradigm switch โ large-scale agentic trajectory fine-tuning plus tool-calling reinforcement learning. The second is harness incompatibility: a 0% score usually means the evaluation harness failed to run the old model at all, not that the model genuinely scored zero.
Both explanations carry the same warning for crypto builders. If the capability is new rather than improved, then every protocol that assumed a small model could drive an autonomous agent was wrong until this generation โ and the 0% figure is the confession. The prior small model had no terminal-operating capability. None. The entire "AI agent on a small model" category was, in part, a product of a number that did not exist.
Even at 39.2%, the ceiling is visible. The same-generation large model scored 70.6% on the same benchmark โ a 44% relative gap. The small model is not a cheap substitute for the large model on long-horizon agentic work. It is a different tool with a structurally lower ceiling. A protocol that routes autonomous capital decisions through a small model to save on inference is not saving. It is buying a lower success rate and booking the difference as margin. On a trading strategy, a lower success rate is not a cost line; it is a variance line, and variance compounds.
Version numbers are a narrative device
OSWorld 2.1. Terminal-Bench 4. Neither version existed at my knowledge cutoff. Both are new. New versions mean prior scores are not comparable.
When a vendor places a new-version score next to a competitor's score, the comparison is only valid if both ran the same harness, the same prompt template, and the same tool set. The release does not state that it did. It does not state who ran the evaluations. In every prior case of self-reported benchmarks I have audited, the gap between self-reported and third-party-verified scores has been material โ sometimes a full capability tier.
There is no LMArena cross-check in the release. No independent analysis number. No reproduction. A benchmark score without a harness specification is a marketing asset, not a measurement. For a protocol deciding whether to trust an agent with capital, an unverified benchmark is worse than no benchmark, because it manufactures false confidence at exactly the moment caution is required.
The price is not a price cut; it is a coordinated floor
One detail deserves more attention than it received. The new pricing was reported as matching, not undercutting, a competitor's small-model tier. The two landed on the same point.
Two competitors arriving at the same price independently is not competition. It is oligopoly behavior. When the two largest labs land on the same number, the number is a floor, not a market price. The competitive axis has moved from price to capability โ rational for the vendors, corrosive for the buyers. A floor means the cost of inference will not fall further through price competition. The only remaining lever is consumption โ and consumption, as established, is rising.
This is the structural tension the sheet exposes. Price is falling. Consumption is rising. Net cost is falling slower than advertised. And the floor means the price lever is exhausted. Everything downstream โ every token underwritten on cheap inference โ inherits the tension.
What this means for the on-chain agent stack
Pull the threads. An on-chain agent protocol has a cost line and a revenue line. The revenue line is denominated in the token it issues or the fees it captures. The cost line is denominated in inference. If the cost line is understated โ by the 1.6x factor above, by the unpinned effort variable, by the unverified capability ceiling โ then the protocol's margin is overstated, and the token that capitalizes that margin is overstated by the same multiple.
I have run this arithmetic on a protocol before. During the AI-agent wallet audit, the team's revenue projection assumed a fixed inference cost per decision and a fixed success rate. Both assumptions were wrong. The cost was higher because consumption was higher. The success rate was lower because the small model's ceiling was lower than the demo suggested. Neither error was disclosed, because neither number was measured. They were quoted from a vendor sheet. The token was priced against a sheet, not a system.
This is the through-line. The crypto industry has learned, painfully, that unbacked yield is a liability. It has not yet learned that unverified inference cost is the same liability in a new denomination. The mechanism is identical: a number is asserted, the number is believed, capital is allocated against the number, and the number turns out to have been an advertisement.
What the bulls got right
Here is the part the bears, and I am one, tend to understate. The agentic capability jump is real, whatever its cause. A small model that can operate a terminal at 39.2% and a desktop environment at 48.9% is a genuinely new building block. Eighteen months ago, the small-model tier could not run an agent loop at all. The 0% score is not only a confession; it is a baseline. The distance traveled from that baseline is large, and it is real.
The bulls are also right that price is falling. A 75% net reduction is a 75% reduction. The error is not in the direction; it is in the magnitude and the certainty. A protocol that plans for a 75% reduction, pins its effort setting, and verifies its model's ceiling on its own harness has a defensible cost model. That protocol exists. It is simply rare.
Where the bulls are wrong is in treating a vendor benchmark as a specification and a vendor price as a cost. Trust is a variable; proof is a constant. The capability is real. The economics are real. The reporting of both is not yet auditable, and until it is, the bull case rests on a sheet that has never been reconciled against a running system.
Takeaway
The next generation of these sheets will report lower prices again. Watch the consumption line. If token-per-task keeps rising faster than price-per-token falls, the net cost curve bends upward no matter what the table says, and every on-chain agent underwritten on a price sheet inherits the error.
The question for builders is not whether the model is capable. It is whether the cost model reconciles. Demand the harness. Demand the consumption figure. Demand the effort setting in writing. The sheet is not the system, and it never was.