Teach Token-Budget Literacy Before Prompting

Teaching token-budget literacy before focusing on prompting is essential for the next generation of AI builders. The current AI curriculum predominantly emphasizes prompting, while neglecting the cost associated with each interaction. As inference prices have dropped significantly in recent years, the focus should shift towards teaching students to estimate, measure, and reduce the cost of AI interactions before and after execution.

To develop cost literacy, students should learn three key habits: estimation, measurement, and reduction. Estimation involves predicting the input and output token count and associated costs before running a prompt. Measurement requires checking the actual token usage against the initial estimate after execution. Reduction entails identifying the most cost-effective prompt or architecture that still produces the desired output.

Concrete pricing data demonstrates the vast cost difference between models. For instance, Anthropic's frontier-tier model, Claude Opus, charges $5 per input and $25 per output, while its mid-tier model, Claude Sonnet, costs $2 to $3 per input and $10 to $15 per output. Google's Gemini 2.5 Flash-Lite efficient model is priced at $0.10 per input and $0.40 per output. The disparity between the highest and lowest rates can be as high as 36 times for input and 89 times for output, depending on the model and provider.

Moreover, there are additional cost-saving measures students should be aware of. Anthropic offers a cache hit discount of up to 90% for repeated context, and batch processing can reduce costs by 50% for both input and output. These habits—estimate, measure, and reduce—should be taught in a single 50-minute class using a current pricing sheet and calculator.

Students can apply these practices by estimating cost for a given task across different models, then running the task to compare real costs and quality. Finally, they should re-run the task using the cheapest model with additional optimization techniques like tighter prompts and caching to observe the reduction in cost.

By integrating token-budget literacy into AI education, students will not only learn to build cheaper AI systems but also create more sustainable solutions, as the same efficiency gains that lowered inference costs also reduced carbon emissions. Teaching these skills early on ensures that future AI builders prioritize efficiency and cost-effectiveness from the start, ultimately benefitting both their projects and the environment.