The report, published on June 26, highlights how different AI tasks demand vastly different amounts of compute, with app and website creation, code debugging, spreadsheet analysis and document generation emerging as the most resource-intensive workloads.
The findings also underscore a growing divide in the economics of generative AI, where increasingly complex tasks require substantially more computing power — and higher API costs — than simpler requests.
Also read: China, India see large companies lose market cap share, hinting at lag in AI raceWhat are AI tokens?
AI models do not read or write in words or sentences. Instead, they process text as tokens—small chunks of text that can be whole words, parts of words, punctuation marks or numbers. For example, the sentence “How are you?” would be broken into several tokens before being processed by the model.
Every prompt a user enters, and every response the AI generates, consumes tokens, making tokens the basic unit used to measure computing usage and API pricing.
Why do some AI tasks cost more than others?
Anthropic measures AI compute in tokens, and since developers are billed on a per-token basis, conversations that consume more tokens are also more expensive to run.
The study found that app and website creation typically uses around three to four times the tokens of a median Claude conversation. Code debugging, data and spreadsheet analysis, and document generation also rank among the most computationally intensive tasks. At the other end of the spectrum, generating an explanation requires roughly one-fifth of the median token count.
| AI task | Token use vs median conversation (Approx) | Relative API cost |
| App/website creation | 3–4x | Highest |
| code debugging | 2–3x | very high |
| Data/spreadsheet analysis | 2x | high |
| Document/report | 2x | high |
| Scripts/snippets | 1.5–2x | above average |
| Plans/Strategy | 1.5x | above average |
| Analysis/summary | Around median | moderate |
| marketing content | below median | lower |
| Email drafting | 0.5x | low |
| Explanation/answer | 0.2x | lowest |
The relative API cost is based on Anthropic’s token distribution and API pricing. Actual costs vary depending on the Claude model used.How much does Anthropic charge?
Anthropic currently charges API developers per million tokens processed, with Claude Sonnet priced at $3 per million input tokens and $15 per million output tokens, while Claude Opus costs $5 and $25 per million input and output tokens, respectively. As a result, more token-intensive tasks can cost several times more to execute than simpler AI requests.
The report also found that higher-value work tends to consume more compute. Conversations associated with higher-paying occupations used roughly twice as many tokens as those linked to lower-paying jobs, with around 44% of the difference explained by the type of output users were creating.
In other words, asking an AI model to explain a concept is relatively inexpensive. Asking it to build an application, debug code or analyze large datasets is a far more compute-intensive and costly exercise, illustrating how the economics of AI vary dramatically depending on the task.
Read more: Why the ₹15,000 smartphone is disappearing
