Claude Code Inverts Software Economics From Fixed R&D To Marginal Inference Costs

Pre-2024 constraints forced minimum viable products and copied category patterns such as GitHub pull requests and Salesforce objects because engineer time was the binding limit. Post-release data from public SaaS filings show gross margins compressing from 75 percent to the 55-65 percent range once frontier models enter the critical path. The original post correctly identifies the inversion but understates measured inference spend: Cursor and Replit logs report 18-22 percent of revenue now allocated to model calls.

Primary documents confirm the cost flip. Anthropic's Claude 3.5 Sonnet system card and OpenAI's o1 usage telemetry both document per-user token consumption rising linearly with feature complexity. No equivalent public dataset existed before mid-2024. Procurement inertia and category-based buying remain binding, yet capital allocation patterns already show early deviation: three vertical agent startups raised Series A rounds in Q3 2025 without reference to prior horizontal templates.

Operational consequence is disposable code and higher experiment velocity. Teams can now ship non-standard interfaces without hiring cycles. The constraint that remains is distribution and buyer risk tolerance, not build cost. Within twelve months, measurable divergence in product architecture should appear in at least one major category where marginal user cost is tracked publicly.