Haueis and Colaço argue that neural metabolism should be treated as a constitutive constraint on cognitive models to assess their biological plausibility and, drawing on extensive neuroscientific evidence, propose ways to evaluate and generate such models. We agree that metabolism shapes neural computations and that ignoring it risks producing biologically implausible models of cognition. However, we identify two limitations in their framework. First, although the authors acknowledge interactions between the brain and the body, their account remains largely neurocentric, treating the brain in relative isolation from the organism in which it is embedded. Second, metabolism is viewed primarily as a constraint that limits neural computations underlying cognition. We argue instead that cognition is not simply limited by the metabolic constraints of the brain, but that cognition actually emerged to serve the metabolic demands of the whole organism. Through such extension and a reorientation of Haueis and Colaço’s framework, we reframe the problem of cognitive modeling from how metabolism constrains neural computation and cognition to how cognition functions to meet the metabolic demands of the whole organism in the service of its survival imperative.

In animals, the nervous system is not metabolically autonomous: its energy supply is regulated at the level of the entire organism, where multiple physiological systems compete for limited resources (Barrett & Simmons, Reference Barrett and Simmons2015). Thus, the architecture of the nervous system, and that of cognition and behavior, has emerged as a solution to the problem of controlling a specific body in interaction with its environment (Chiel & Beer, Reference Chiel and Beer1997; Cisek, Reference Cisek2019, Reference Cisek2022). The organization of the octopus’ nervous system provides an illustration of this. With an invertebrate body, maintaining a point-to-point (somatotopic) brain representation of different body parts would be metabolically prohibitive. Instead, the octopus relies on a highly decentralized nervous system, with approximately two-thirds of its neurons located in its arms. This organization enables the use of peripheral sensory tagging for motor control: when an arm contacts a stimulus, the sensory input does not merely signal its presence but specifies the location for action, effectively serving as an address for motor commands. The brain just sends an en passant motor signal along the arm, which is only executed where it coincides with local sensory input. In this way, sensory input functions as a temporary, action-guiding map without the need for central somatotopic representations (Hochner et al., Reference Hochner, Zullo, Shomrat, Levy and Nesher2023). Thus, in the octopus, the implementation of motor control reflects a solution to organism-level metabolic constraints, rather than solely neural ones.

Crucially, the same metabolic principles extend to higher-level cognitive functions such as reinforcement learning and decision-making, where metabolism does not merely constrain but actively generates behavioral policies, shaping which actions are selected and how they are executed. Organisms continuously allocate limited energy across competing needs, such as whether to forage or rest, explore or exploit. These trade-offs are governed by the current metabolic state of the body, which cannot be inferred from the cost of neural activity alone. Hunger or satiation systematically shape value computation, action selection, and execution (Juechems & Summerfield, Reference Juechems and Summerfield2019; Keramati & Gutkin, Reference Keramati and Gutkin2014; Yoshida et al., Reference Yoshida, Daikoku, Nagai and Kuniyoshi2024). For example, animals that have learned to avoid a salt-associated cue can immediately reverse this preference when sodium-depleted, approaching the stimulus without any new learning (Robinson & Berridge, Reference Robinson and Berridge2013). While the stimulus remains unchanged, the organism’s internal state has entirely restructured its value and the cost–benefit logic of the learned associations. Similarly, metabolism dynamically regulates the energetic investment of action itself. Animals invest additional energy to move faster only when the anticipated benefit justifies the metabolic cost (Shadmehr & Ahmed, Reference Shadmehr and Ahmed2020; Starita et al., Reference Starita, Garofalo, Dalbagno, Degni and di Pellegrino2022; Summerside, Shadmehr, & Ahmed, Reference Summerside, Shadmehr and Ahmed2018). This calculation shifts dynamically with the organism’s physiological state, such that, in conditions of metabolic depletion or threat, the neural and physical cost of a vigorous action is discounted by the gain of physiological restoration (Garofalo et al., Reference Garofalo, Finotti, Starita, Bouchard, Fecteau, Boyle, Northoff, Barbey, Fregni, Jahanshahi, Pascual-Leone and Sahakian2024; Keramati & Gutkin, Reference Keramati and Gutkin2014). These examples suggest that the computations underlying decision-making are organized in the service of the metabolic needs of the body rather than being merely constrained by it. Considering only the energetic costs of neural computations risks mischaracterizing the function and organization of cognition, as it may overlook how the fluctuating metabolic needs of the body define the very purpose and objective of the neural computation.

Our framework also reshapes the comparison between biological and artificial systems. The relevant question is no longer whether artificial systems match the energy constraint of the brain, but whether their architectures incorporate analogous principles of whole-body metabolic regulation. Standard computational models of reinforcement learning generally treat goals and values as externally fixed, with rewards and costs specified in advance. However, the above empirical evidence shows that this is not how learning and decision-making operate in biological agents. Homeostatic reinforcement learning provides a formal computational model that more closely mimic biological cognition, replacing abstract reward maximization with (metabolic) drive reduction (Keramati & Gutkin, Reference Keramati and Gutkin2014). In such models, the value of an action is dynamically scaled by the urgency of a physiological deficit. These systems show that intelligent, adaptive behavior can emerge without externally specified rewards, for example, by learning to forage for nutrients solely by monitoring and optimizing the maintenance of internal metabolic set points (Yoshida et al., Reference Yoshida, Daikoku, Nagai and Kuniyoshi2024). Embodied artificial systems thus show that metabolism can function as a genuinely generative architectural foundation also for artificial cognition.

In conclusion, Haueis and Colaço have made a compelling case that metabolism belongs at the center of cognitive modeling. We have argued that this center cannot hold at the level of the neuron or the brain alone; it must expand to encompass the organism as a whole, the body whose survival imperatives give neural computation and cognition both its constraints and its meaning.