The Substitution Wave in AI

The article exposes a growing tension in AI: frontier model prices are rising, foundation labs are moving up the stack into applications, and open-source models have crossed the good-enough threshold for most use cases. The natural response from buyers is substitution — actively routing prompts to cheaper models or switching entirely. This creates a structural shift where the old assumption of always using the best closed model no longer holds.

Concrete examples show how companies are executing this substitution. Coinbase routes prompts to cheaper models and keeps costs flat while token usage grows exponentially. Lindy switched 100% of traffic to DeepSeek v4, saving millions and seeing performance increases. Harvey found that SFT on Kimi 2.6 beat Opus on a legal benchmark while costing 11x less ($84 vs $954 across 100 tasks). Cursor went further by post-training Kimi K2.5 into its own Composer model, claiming up to 10x more efficiency than similarly capable models.

The takeaway is that buyers don’t pocket the savings — they reinvest in more intelligence, keeping costs flat while usage grows exponentially. Closed models get more expensive at the frontier; open models get cheaper at parity. The strategic choice is which cost slope fits your unit economics. The substitution wave is not a temporary hack but a fundamental shift in how AI infrastructure is built and consumed.