Frontier AI Labs Hold Premium Pricing Ground Even as Open Source Models Flood Enterprise Pipelines

A provocative new framework for understanding the AI economy emerged this week when Decagon CEO Jesse Zhang published a blog post challenging the conventional wisdom that open source AI models are eating into frontier labs' business. Zhang's central argument is that frontier models like Anthropic's Claude and OpenAI's GPT series are not competing with open source alternatives so much as passing the baton to them — expensive, cutting-edge models prove out new use cases first, and then those proven workflows migrate to cheaper open source alternatives as the technology matures and the use cases become better understood. As fast as those mature deployments move downmarket, entirely new use cases keep entering the pipeline and anchoring fresh spend on frontier models.

The data backing up this thesis is surprisingly accessible. Vercel's AI gateway dashboard reveals that DeepSeek has surged to lead all models by raw token volume in just the past week, now handling roughly a third of all tokens flowing through the platform. Meanwhile, Z.ai's GLM-5.2 model has muscled its way into fourth place over the same period, signaling a real shift in how developers are routing their workloads. Yet scrolling down to spending figures tells a different story entirely: Anthropic still accounts for more than half of total AI spend across the Vercel platform, a dominance that has dipped only slightly despite the company's own recent price increases.

OpenRouter, which serves a broader and somewhat less enterprise-focused user base, reinforces the pattern. DeepSeek V4 Flash leads on sheer token volume at 5.3 trillion tokens weekly, while the frontier model Opus 4.8 handles just over 2 trillion. However, OpenRouter's own data shows Opus 4.8 carries an average token cost approximately 23 times higher than V4 Flash — $1.37 per million tokens versus a mere six cents — meaning Opus is almost certainly capturing a dominant share of total monetary spend on the platform despite trailing badly on volume. Adding further complexity, Nvidia's new Nemotron model has arrived with significant enterprise momentum thanks to Nvidia's deep industry relationships and the model's noted adaptability across tasks.

Zhang's blog post does not offer exhaustive datasets of its own, but the independently observable market signals lend credibility to his lifecycle framing. One interpretation is that the total universe of AI-addressable tasks is expanding so rapidly that frontier labs can sustain their revenue simply by perpetually owning the experimental, early-stage end of deployment — what Zhang summarizes as, "The frontier labs will keep owning discovery. Open source will increasingly own production." An alternative explanation is that many enterprise use cases are sufficiently complex or sensitive that they cannot be fully handed off to lighter alternatives regardless of cost pressure.

This dynamic represents a notable shift from predictions made as recently as late 2024, when some analysts warned that foundation model providers could become commodity input suppliers while application-layer companies captured most of the value. That scenario has only partially materialized: while vertical AI applications have indeed migrated toward lighter models and "GPT wrapper" startups have stabilized economically, frontier providers have successfully held their position at the premium end of the pricing spectrum. Whether this two-tiered structure becomes a durable feature of the AI economy or merely a transitional moment remains an open and consequential question.