How Should Researchers Report the Computational Cost of AI-Assisted Research?

What Should a Research Paper Reveal About the Cost of Using AI?

A research paper reports that a language model classified 20,000 documents. The authors identify the model and describe the prompts, but provide no information about API usage, processing time, computational resources, or the expense of repeating the analysis.

Another researcher wants to reproduce the workflow. How much computing capacity would be needed? Would the analysis cost a few dollars or several thousand? Were unsuccessful runs and validation experiments included in the reported expenditure?

These questions are particularly relevant when AI is used extensively or when access to computational resources affects who can independently evaluate the findings. Transparent cost reporting helps readers understand the resources behind a method, although the appropriate level of detail depends on the study.

Report Resource Consumption, Financial Costs, and the Assumptions Behind Them

What Computational Cost Reporting Should Include

Why computational cost belongs in research methods

Methods sections explain how research findings were produced. When AI processing is an essential part of the methodology, the resources required to execute that processing may be relevant to evaluating and reproducing the study.

For conventional computational research, reporting software, hardware, and computational requirements can help other researchers understand the conditions under which analyses were conducted. Similar reasoning applies to AI-assisted workflows, although different deployment arrangements require different forms of documentation.

A cloud-based language model may generate financial charges according to usage. A locally deployed model may consume GPU time, electricity, and storage without producing a commercial API invoice. A university-hosted model may appear free to the individual researcher because the institution pays for the infrastructure.

These arrangements are not directly comparable using the amount personally paid by the researcher.

The purpose of reporting is not to demonstrate that an expensive method is scientifically superior. It is to disclose enough information to understand the computational demands and practical accessibility of the reported workflow.

Distinguish computational consumption from financial expenditure

Computational consumption describes the resources used to perform a task. Financial expenditure describes the monetary amount paid or attributed to that activity.

Both can be informative, but neither should be presented as a substitute for the other. For example, a research team with donated cloud credits may report zero out-of-pocket computing charges while still documenting substantial GPU usage.

Similarly, an institution may already own the required hardware. Its researchers might incur no new equipment purchase, but the study still consumes machine time and other resources.

What to report for API-based AI research

Application programming interfaces, or APIs, allow research software to send requests to hosted AI systems. Many commercial providers charge according to usage, although billing structures vary.

For a substantial API-based study, useful reporting details may include:

• The provider and exact model identifier, including a version or dated snapshot when available.
• The period during which processing occurred.
• The number of requests or processed research units.
• Total input and output tokens when the service exposes reliable usage records.
• Any relevant cached-input, batch-processing, reasoning-token, or additional service charges.
• Actual charges, applicable currency, and the basis of the cost calculation.
• Whether reported usage includes development, validation, failed requests, and repeated runs.

A token is a unit used by a model's text-processing system. Token counts are not interchangeable with word counts, and different models may tokenize the same text differently.

Where a provider exposes separate billable categories, researchers should use those categories rather than assuming every token has the same price.

For example, official API documentation from OpenAI and Anthropic describes provider-specific usage and pricing arrangements. These arrangements can change, so researchers should preserve the applicable pricing information and verify it against the official documentation when reporting costs.

What to report when AI models run locally

Locally executed AI workflows require different measurements because there may be no per-request commercial charge.

Relevant details may include the processor or accelerator model, number of devices, approximate accelerator-hours, CPU resources, memory capacity, model precision, software framework, and total processing time.

Researchers should distinguish allocated hardware time from actual measured utilization when that difference matters. Reserving eight GPUs for an hour does not necessarily mean all eight devices operated at full computational capacity throughout the period.

The hardware requirements for running AI models locally depend on model architecture, parameter count, numerical precision, context length, workload, and implementation.

For reproducibility, documenting the model weights or release identifier and inference software may be as important as reporting the hardware.

Should researchers report GPU-hours?

GPU-hours provide a simple measure of accelerator allocation over time. One GPU used for four hours represents four GPU-hours. Four GPUs allocated for the same four-hour period represent sixteen GPU-hours.

This calculation is useful for documenting resource allocation. It does not establish actual floating-point operations, energy consumption, or hardware utilization. Twelve GPU-hours on different accelerator models may represent substantially different computational capabilities and electricity demands.

Researchers should identify the hardware and measurement basis rather than reporting GPU-hours without context.

How to calculate API processing costs

For a service with separate input and output token prices, a basic cost estimate can be calculated from the billable token counts.

The result represents a hypothetical calculation, not a quotation from an actual provider. It also excludes other charges, taxes, discounts, and processing categories that may apply.

For actual reporting, provider invoices or recorded billing data are preferable to estimates when they can be accurately attributed to the research activity.

Where a project uses a shared account, the researchers should explain how study-specific charges were identified or allocated.

Should failed requests and repeated experiments count?

They should be considered when defining the reporting boundary.

AI research often involves preliminary prompt testing, failed requests, repeated outputs, model comparisons, and validation experiments. Some activities may consume billable resources even when they do not produce results included in the final analysis.

Reporting only the successful final processing run can substantially understate the resources required to develop the reported workflow.

However, the total expenditure of a research project and the cost of reproducing its finalized analysis are different quantities.

Separating these categories is often more informative than reporting a single unexplained amount.

How should subscription costs be reported?

Flat-rate subscriptions create an allocation problem because the same account may support several research projects, teaching activities, or personal tasks.

A researcher should not automatically attribute the entire subscription fee to one study when the service was also used for unrelated purposes.

Possible approaches include reporting the actual subscription expenditure during the study period, explaining an allocation method, or stating that study-specific usage could not be separated reliably.

For institution-wide licenses, the research team may not know the negotiated price. In that case, reporting the service tier, usage volume where available, and the limitation on financial attribution may be more defensible than inventing a monetary estimate.

A subscription fee should not be converted into a supposed per-token price unless the calculation has a clear and appropriate basis.

What about the cost of human validation?

Human verification is not computational consumption, but it can be a substantial part of the total cost of AI-assisted research.

For example, an inexpensive classification model may generate outputs that require extensive manual correction. A more expensive model may reduce that workload, although such savings must be demonstrated rather than assumed.

Researchers can report validation labor separately, using recorded person-hours and a clearly identified valuation method when financial estimates are appropriate.

Combining human labor with computational expenditure without explanation can obscure the relative contribution of each.

The distinction is particularly important when cost is considered in comparing AI models.

Why reporting dates and pricing assumptions matters

Commercial AI prices, available models, subscription conditions, and billing categories may change during or after a study.

Therefore, a cost estimate should identify the relevant pricing period and whether it represents actual charges, a published list price, or a modeled estimate.

Currency conversions should state the exchange-rate basis when material. Researchers should also disclose whether credits, discounts, subsidies, or institutional agreements affected the reported expenditure.

A study conducted using donated credits may have a low financial cost to the authors without being inexpensive for an independent researcher purchasing the same service.

This distinction matters because unequal access to paid AI services can influence the feasibility of reproducing computationally intensive research.

How much reporting detail is enough?

Not every AI interaction requires a separate cost table. If AI was used only for minor, non-analytical assistance, extensive computational accounting may contribute little to understanding the study.

More detailed reporting is appropriate when AI processing is central to the method, computational expenditure is substantial, resource limits influence the design, or the analysis is intended to be reproduced independently.

Researchers should also check the target journal's current reporting requirements. There is no universal requirement that every research article disclose identical AI cost metrics.

For computationally intensive machine-learning research, the NeurIPS Paper Checklist provides an example of a venue-specific approach to documenting computational experiments and reproducibility-related information. Its requirements should not be generalized automatically to all disciplines or journals.

The principle is proportional transparency: report enough to make the computational method understandable and its resource requirements reasonably assessable.

Can computational cost reporting support environmental assessment?

Yes, but only within the limits of the measurements available.

GPU-hours, token counts, and API charges may help characterize the scale of an AI workload. None of these measures directly establishes electricity consumption or greenhouse-gas emissions.

Environmental estimates may require additional information about measured power, hardware, infrastructure efficiency, geographic electricity characteristics, and the boundaries of the assessment.

Researchers should not convert financial expenditure into environmental impact using an unsupported universal conversion factor.

The related question of reporting computational resource use in large-scale AI research deserves particular attention when workload size makes energy and infrastructure demands material.

Reporting the Cost of an AI-Assisted Literature Classification Study

The final execution cost is $15.00, while the complete API expenditure attributable to the illustrated development, validation, and execution stages is $21.00.

The researchers could report the computational resources as follows:

In an actual manuscript, the researchers should replace the hypothetical provider description and rates with their recorded model identifier, processing dates, billing data, and applicable conditions.

The calculation also illustrates an important distinction: the $15.00 final execution cost may help estimate what another team would spend to repeat the finalized processing stage, whereas the $21.00 total better represents the computational expenditure across the reported development and validation workflow.

Neither amount establishes that the classifications were accurate. Model performance and human validation results must be reported separately.

Common Errors in Reporting AI Research Costs

How to Build Computational Cost Reporting Into Your Research Workflow

Computational cost reporting is easiest when resource consumption is recorded during the study rather than reconstructed after publication.

Before processing begins, establish a resource log. Record the model identifier, processing date, task, number of requests or processed items, and available usage measurements. For local models, record the hardware configuration and allocated processing time.

Separate exploratory development from the finalized workflow. This distinction allows readers to understand both the resources required to establish the method and the cost of repeating its final execution.

For larger projects, a supplementary resource table may be more readable than placing every usage figure in the main methods section.

When selecting models, the reporting framework can also help compare lower-cost and open AI alternatives using documented resource requirements rather than advertised prices alone.

Ultimately, computational reporting should enable another researcher to understand what was consumed, what was paid, what was subsidized, and what would plausibly be required to repeat the work.

Before Submitting an AI-Assisted Research Manuscript

Questions About Reporting AI Computational Costs

Transparent Cost Reporting Makes AI Research Easier to Evaluate and Repeat

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.