H&C convincingly point out that metabolic considerations (MCs) have been unduly ignored in part because multiple realizability and medium independence theses have led cognitive scientists to view them as inconsequential to characterizing cognitive processes, and because the general view of the brain as energy-efficient is sometimes taken to license their neglect. The authors make a valuable contribution by showing that these are bad reasons to ignore MCs and that such facts should be included in the general stockpile of evidence that constrains and indicates models of cognition.

However, one might wonder whether MCs are distinctive in their potential to inform cognitive theory. H&C’s complaint that they are often invoked too generically is used to motivate the need for a systematic methodology. But if MCs are just another sort of biological detail, then there is no reason to suppose they require their own methodology distinct from one that utilizes biological details more generally.

Here is why they might be distinctive. All cognitive processes consume energy, so ATP serves as a common currency that every cognitive process must spend, forming a particularly informative bridge between low-level biological facts and higher-level theories of cognition. By contrast, the fact that, say, a synapse typically releases a single kind of neurotransmitter can constrain plausible cognitive models but can hardly constrain high-level theories, since it bears no direct connection to them. A Bayesian model pitched at the highest level of analysis, on the other hand, can still be ruled out if it requires more information transfer than ATP budgets could plausibly support. Because MCs bear on efficiency, the thought goes, they might constrain cognitive theories in unique ways, warranting a specialized methodology.

Interestingly, it is not clear that the examples H&C focus on actually illustrate MCs playing this kind of distinctive role. In the case of the core evaluative example, the linear cascade model, the point is simply that synapses at equilibrium are dead, and dead synapses cannot store memories. The constraint that cells be living applies equally to non-cognitive biological processes.

The supplementary example of the leaky integrate-and-fire models being assessed as less plausible is not due to considerations of resource efficiency but simply that EDLIF is more empirically accurate when it correctly models the Na⁺/K⁺ pump’s activity as dependent on ATP. While modeling ATP’s causal effect on other variables is in some sense ipso facto about energy use, there is no analysis of ATP per bit or the like; there is no bridge to cognition of the kind that would make MCs distinctive. The model is already cashed out at the biological-detail level and is being de-idealized by including a further causal variable, as could be done with any biological detail.

As for their core generative example, the efficiency-tradeoff hypothesis does not generate a novel cognitive model. Instead, it takes a pre-existing cognitive model (predictive processing) with already existent specified implementational details (thin interneuron axons transmitting error signals) and uses MCs merely to rationalize this design choice.

Part of the reason why it is difficult to use MCs in a non-generic way may have to do with the nature of efficiency considerations in general. Consider the oft-remarked point about the brain’s incredible efficiency. Whereas frontier LLMs can draw up to 1000 watts during inference, brains run on around 20 watts. This seems to close the question of which is more efficient, but the story is complicated once one asks what one is getting for such a wattage. ChatGPT-5 can, according to estimates, use around 34 Wh on reasoning-heavy tasks, which is roughly what the human brain uses in two hours (Jegham et al., Reference Jegham, Abdelatti, Koh, Elmoubarki and Hendawi2025). If one supposes that it would take a human more than two hours to produce equivalent output, then ChatGPT would be, in fact, more efficient than the human brain on that task. The point is similar to one that H&C rightly note: ATP/second is not the correct measure of efficiency. ATP/bit, as they suggest, is better. Shannon information, however, is famously semantically blind. All ChatGPT outputs of the same token length, produced with similar inference times, will have a similar electricity per bit, yet one output might take a human significantly less than two hours to replicate, and another significantly more.

The lesson to draw out here is that to assess resource efficiency, one needs to know the goal of cognition. Without knowing what the task is, what is being accomplished, and how well it’s being done, it’s not possible to assess whether ATP is being used wisely or wastefully. Icard (Reference Icard2023) makes the point that to know the value of a resource, and whether it’s well spent, you want to know what else you could have spent those resources on.

This requirement – evaluating efficiency relative to task-specific opportunity costs – is exactly what resource rational analysis (RRA) systematizes. RRA begins by establishing a task-environment and deriving an optimal cognitive model. Discrepancies between predicted and actual behavior indicate psychological constraints, which are incorporated to derive a new model that is resource-optimal relative to bounds. This new model provides a baseline from which further discrepancies can be observed, and the methodology iterates.

Such a methodology provides exactly what H&C want, despite their claim that RRA lacks a unifying account of metabolic and informational constraints. In fact, RRA translates ATP and MCs, in general, into a common currency by treating them as limited resources that constitute task-indexed constraints, which can then be traded off against other kinds of constraints. An explicit step is deriving optimal models from the set of identified constraints, which provides precisely the generative use that H&C claim is missing from RRA. This is the advantage of the top-down approach that RRA offers by initially deriving optimality directly from task-environments and working its way down, whereas a bottom-up approach that starts with MCs can only employ them as generic biological details. If one wants a systematic methodology that treats MCs as a distinctive kind of biological detail, one should do RRA.