Why the rise of open source AI isn’t hurting Anthropic … yet
TL;DR
TechCrunch builds on a thesis from Decagon CEO Jesse Zhang: companies often test new AI use cases with expensive frontier models, then move mature workflows to cheaper open source models. Vercel data shows DeepSeek handling more than a third of token volume, with Z.ai also ranking high. But Anthropic still takes more than half of AI spend on the platform. OpenRouter shows the same split: DeepSeek V4 Flash processes far more tokens than Opus 4.8, while Opus costs roughly 23 times more per million tokens.
Nauti's Take
The clean read: open source is winning a lot of production work, while Anthropic keeps monetizing uncertainty. That is a comfortable setup for frontier labs as long as customers pay for the first messy phase.
The catch is that better evaluations, routing, and fine-tuning routines will make the premium share harder to defend. Anthropic is not safe because open source is weak; it is safe because the market keeps creating fresh hard problems.
Briefingshow
Open source is not automatically eating Anthropic because usage and revenue are splitting apart. For teams, model strategy becomes two-stage. Use strong models to find out whether a workflow works at all, then cut costs once the process is stable.
Token volume alone misses where the money stays.
Sources