What happened
Decagon CEO Jesse Zhang theorised a two-phase AI model lifecycle: frontier models for discovery, open-source for production. Vercel data shows DeepSeek processing over one-third of tokens, yet Anthropic maintains over half of overall AI spend. OpenRouter reports DeepSeek V4 Flash handles 5.3 trillion tokens weekly versus Opus 4.8's 2 trillion, but Opus 4.8 costs 23 times more per million tokens, ensuring its spend dominance. Nvidia's Nemotron is also poised to enter this market.
Why it matters
AI procurement teams face a two-tiered model economy where frontier models capture premium spend for discovery, while open-source models handle production. This dynamic means initial project costs remain high for proving new use cases, despite the availability of cheaper, high-volume open-source alternatives for mature deployments. Platform engineers must plan for model transitions, recognising that the cost curve for advanced capabilities remains steep, a trend consistent with recent reports of soaring AI costs forcing enterprise re-evaluation.
