Enterprise AI buyers are substituting closed frontier models for cheaper open-source alternatives at scale. Three forces are driving this: foundation labs moving up the stack into applications, frontier model prices rising for the smartest closed models, and open-source models crossing the good-enough threshold for most production use cases.
The case studies are concrete. Lindy switched 100% of its traffic to DeepSeek v4, churning from Anthropic, saving millions while reporting performance gains. Harvey ran a 100-task legal agent benchmark where a fine-tuned Kimi 2.6 beat Claude Opus at 15% vs 14% all-pass rate, at $84 versus $954 per 100 tasks, roughly 11x cheaper. Cursor post-trained Kimi K2.5 into its own production model, Composer 2.5, claiming 10x efficiency gains over comparable closed models. Coinbase is routing prompts to cheaper models dynamically, holding costs flat while token usage grows exponentially.
That last Coinbase detail is the one worth reading the original for. Cost savings are not being pocketed. They are being reinvested in more inference. The real question the piece raises, without fully answering, is what happens to closed model incumbents when open-source parity holds and buyers have a clear cheaper slope for their unit economics.
[READ ORIGINAL →]