A company moves a workflow from an expensive frontier model to a smaller model that tries several times, consults tools and checks its own answer. The cloud bill falls. The product team celebrates an efficiency win. But what happened to the electricity used by the complete job? That is a different question, and it is becoming more urgent as AI products shift from answering prompts to executing multi-step work. A November research preprint supplies an important warning: cost per token, carbon per model call and carbon per successfully finished task are not interchangeable. A system may be cheaper while consuming more computing work than the single response it replaced. It may also use additional computation to avoid expensive failures. The relevant unit is not the model's published price. It is the verified outcome and everything required to produce it.

The study that exposes the accounting problem

Boris Ruf, Adrien de Mailly Nesle and Marcin Detyniecki posted “Toward Sustainable Agentic AI: Carbon Tradeoffs Across Iterative Self-Refinement and Tool-Augmented Coding” on November 9, 2026. Their work, presented at a GREEN AI workshop, examined two experimental settings. In the first, a smaller model revised answers on mathematics tasks across one to five passes and was compared with a larger single-pass baseline. In the second, coding agents equipped with tools operated in isolated software environments on SWE-bench Lite. These are controlled research scenarios, not a comprehensive census of deployed agents. Their value is in making the accounting assumptions visible.

In the mathematics setting, three refinements nearly closed a 16-percentage-point accuracy gap: the remaining difference was within roughly one point, according to the authors. Yet the repeated passes nearly tripled the estimated carbon footprint relative to the cheaper model's single-pass baseline. Under the study's token-based carbon assumptions, that repeated process still remained below the larger model's one-shot estimate. Neither result supports a blanket claim that more reasoning is greener or dirtier. It shows that the answer can change depending on the baseline, the number of loops and the accuracy one expects. A cheap attempt is an input; a correct completed task is an outcome.

The denominator can reverse the winner

The coding-agent experiment makes this distinction sharper. The researchers report task resolution rates of 81.2% for GPT-5.3-codex, 68.9% for Claude Opus 4.6 and 25.0% for GPT-4.1-mini in their sample. A model that produces the fewest estimated emissions per individual call can be a poor choice if it requires many retries or fails most jobs. Conversely, a powerful model might finish in fewer attempts and spare developers hours of debugging, even if each inference is more intensive. These observations do not provide an all-purpose ranking; they establish that performance and carbon must be normalized to comparable work and reported with uncertainty.

The most important caveat is the one the authors explicitly emphasize. Their carbon estimates cover output-token inference only. In the coding study, prompt tokens accounted for more than 99% of recorded token volume, while the carbon attributed to those prompt tokens and to tool execution was excluded. That means the modeled figures are lower bounds on full agentic work, not measurements of total operational emissions. Anyone turning the paper into a precise claim about which current model is the greenest would overstate its evidence. The blind spot is also a useful map of what cloud platforms, model providers and enterprise operators need to measure next.

An agent is a workflow, not a chat response

Imagine an agent tasked with repairing a software bug. It retrieves repository context, reads test output, drafts a patch, runs a container, inspects a failure, asks another model to critique the patch and tries again. A pricing calculator that multiplies output tokens by a published rate misses the repeated reading of long contexts, code-execution machines, storage operations and perhaps even developer time spent intervening. A single request may complete in seconds; another may run for hours. The difference resembles an industrial process with multiple stages more than a traditional web search. It is misleading to label both operations a “query” and then divide their costs by a single request counter.

The same boundary problem appears in environmental reporting. An operator can count energy attributable to a serving GPU but omit cooling and networking, or report data-center electricity without allocating it to actual workflows. Geographic and temporal differences in electricity supply also matter for estimated emissions. The sources used here do not quantify the hourly emissions attributable to the particular agent experiments. The International Energy Agency's 2026 update describes the simultaneous improvement in efficiency for simpler uses and the growing popularity of far more energy-intensive AI applications. Aggregate data-center demand and per-workflow accounting must be analyzed at different scales; one cannot be inferred directly from the other.

Three metrics worth separating

A useful enterprise dashboard would report at least three distinct measures. First is computational demand: prompt and output tokens, number of model invocations, retries, tool calls and duration of execution. Second is quality-adjusted work: accepted tasks, verified outcomes, human escalation and downstream rework. Third is environmental impact: estimated or measured electricity and associated emissions across the relevant infrastructure boundary, with an explicit location, time horizon and uncertainty interval. These figures can be combined for decisions but should not be collapsed into one unexplained sustainability score. A low number in one column may merely transfer work into another.

For example, consider two hypothetical support agents handling 1,000 tickets. One completes 800 correctly on first pass, while another resolves 930 after a series of retrieval, verification and escalation steps. Even if the latter consumes more electricity per attempt, it may reduce repeat customer contacts, manual investigations and customer churn. Whether its overall environmental effect is lower cannot be known without counting those effects consistently. The illustration is not a numerical conclusion from the preprint. It shows why “tokens per ticket” may be useful for finance but insufficient for an evaluation of service quality or whole-system energy consumption.

The rebound effect is an open question

Improved efficiency can also expand demand. If agents become dramatically cheaper to deploy, a business might automate thousands of low-value tasks it would never previously have commissioned. Unit costs fall while aggregate computing and electricity demand rise. Economists would recognize a familiar rebound possibility, but the magnitude for agentic AI is not yet established by the study. The IEA projects substantial growth in data-center electricity use and emphasizes uncertainty in future demand, but global forecasts cannot identify the rebound effect of a single product design. Developers should therefore distinguish efficiency claims about one workload from claims about the industry's eventual total impact.

A practical experimental design for buyers

Before purchasing a supposedly greener agent product, an enterprise can define a repeatable workload basket: short routine tasks, complex cases, tool-heavy jobs and failure cases. It can then compare candidate systems at the same acceptance standard, rather than forcing all of them to the same token budget. Capture model invocations, context lengths, tool compute and retries through instrumentation. Record failed and partially completed jobs. Where electricity measurements are unavailable, show a range using disclosed assumptions rather than replacing missing data with a zero. Make the experiment repeatable by fixing data snapshots and tool versions. Include the cost of rework when a task appears complete but is later rejected.

For teams building their own agents, the fastest improvements may be architectural rather than model-related. Stop loops when evidence is sufficient; cap retry budgets; route straightforward work to narrow deterministic software; cache reusable retrieval results safely; and reserve expensive verification for material risk. But every optimization needs a paired quality check. A shorter prompt that silently omits a necessary document could cut emissions while degrading correctness. Equally, a second reviewer may be worth the energy when a faulty action would impose significant harm. Operational efficiency is an optimization problem with safety, reliability and customer value as constraints—not an unconditional race toward the smallest possible number of calls.

What the new research does not prove

The paper does not directly meter the electricity used in every data center or compute a full life-cycle assessment. It estimates a subset of operational impact under stated assumptions, compares particular models and benchmark tasks, and cautions that omitted costs are consequential. Its headline rankings should not be generalized to other models, new hardware or different coding tasks without replication. The preprint is especially useful precisely because it acknowledges those limits. Research that exposes the size of an accounting gap can be more valuable than a deceptively precise estimate that hides it.

The business case for honest accounting

Model suppliers and agent vendors have an opportunity to compete on verified useful work per unit of cost and energy, not just inference speed or token pricing. Enterprise customers should request auditable workload-level reports and make evaluation terms explicit: what counts as completion, which tool infrastructure is included, and how retries and human intervention are treated. Regulators may eventually seek comparable disclosure, but organizations need not wait for a universal standard to improve their own measurement. The immediate decision is simpler. If an agent is sold as both economical and sustainable, insist that both claims refer to the whole job. A small model is not automatically a small footprint.

For the infrastructure side of this story, read why temporary gas plants can create long-term AI infrastructure lock-in. For the buyer's side, see why cheap tokens alone do not close the AI adoption gap.

Sources and methodology

Experimental figures are attributed to the authors' preprint and should be read with its stated modeling exclusions. IEA figures provide system context, not a measurement of the specific benchmark runs. Enterprise examples, evaluations and recommendations are Nextwith.ai analysis.