Claude Code Development Cost Analysis: Subscription is 2%, Human Judgment is 98%

A developer analyzed the actual cost structure of AI coding tools by building a Baroque music generation system based on the Bach corpus over six months using Claude Code Max 20×. Out of 550 hours of total work time, subscription fees accounted for only 2.0%, with the majority of costs attributed to human listening and refusal processes to compensate for the AI's lack of judgment.

Cost Structure and Efficiency

• Total Cost: Approximately 56,100 CHF (55,000 CHF for personal work time + 1,120 CHF for subscription fees)
• Cost Savings: 12–21x savings compared to estimated manual costs (680,000–1,190,000 CHF)
• Subscription Ratio: At 160 CHF per month, equivalent to 1.6 hours of personal time, it secured efficiency gains of 890–1,620 hours per month
• Key Finding: The low cost of 'being wrong' enabled a 'build-measure-refute-rebuild' iteration cycle over upfront design, driven by reduced failure costs rather than increased typing speed

Output Scale and Metric Criticism

• Code Scale: 148,284 lines of manual SLOC, 113,000 lines of design documentation, and 450 defect records
• Edit Traffic: 3.68 million lines inserted vs. 2.62 million lines deleted (6.3 million total lines edited, with a code rewrite ratio of approximately 3.5x)
• Limitations of Cost-per-line Metrics: This metric is unreliable due to inflation caused by AI-generated volume (e.g., module splitting, documentation synchronization). Costs should be calculated based on capability.

AI Limitations and the Human Role

• Importance of Judgment: Most of the 550 hours were spent on humans listening and refusing results, as Claude cannot hear sound.
• Inability to Parallelize: Batch processing is impossible due to questions depending on previous answers, requiring 4–5 rounds of verification per thread.
• Maintenance Risk: Stopping subscription payments prevents the enforcement of documentation rules, leading to internal decay (doc-drift), where this recurring maintenance cost may exceed the subscription fee.

Quality Assessment

• A blind evaluation by ChatGPT resulted in a score of 25/30 (ABRSM Grade 8 level), but limitations were noted regarding predictability due to mechanical structure and a lack of human rhetoric.