A multinational enterprise reviews AI operating expenses across several standardized workflows. As the Chief Data & AI Officer (CDAO), you observe…

The operational cost of most large language models (LLMs) is directly proportional to the number of tokens processed, which includes both the input prompt and the generated output. Examining "high token consumption per task" is the most direct way to determine if the structure of the AI interaction (i.e., the length and complexity of prompts and expected responses) is driving the cost variance. This metric holistically captures the total workload placed on the model for a given task, making it the primary factor to investigate for cost differences in standardized workflows.

B. This describes a potential cause of inefficient prompts (e.g., user skill), not the direct, measurable behavior of consumption itself.
C. This is incomplete as it only considers the input prompt's length, ignoring the significant cost contribution from the generated output tokens.
D. This behavior would manifest as higher total token consumption across multiple interactions, making it a symptom rather than the core metric to be measured.

1. OpenAI. (2023). Pricing. OpenAI API Documentation. The official pricing model is explicitly defined per 1,000 tokens for both input (prompts) and output (completions), confirming that total token count is the direct driver of cost.

2. Brown, T. B., et al. (2020). Language Models are Few-Shot Learners. Advances in Neural Information Processing Systems, 33, 1877-1901. (Available via NeurIPS proceedings). This seminal paper on GPT-3 establishes the input-output structure where the length of the prompt and completion determines the computational load, which translates directly to cost.

3. Google Cloud. (n.d.). Generative AI on Vertex AI pricing. Google Cloud Documentation. The pricing for models like Gemini is listed per 1,000 characters or tokens for both input and output, reinforcing token consumption as the universal cost metric.