The open-source AI movement has produced genuinely impressive models. DeepSeek now processes over a third of all tokens flowing through Vercel’s AI gateway. Z.ai’s GLM-5.2 sits comfortably in the platform’s top four by volume. These are not hobbyist projects. They are production-grade systems handling real enterprise workloads at massive scale. Given these numbers, it would be easy to assume that frontier labs like Anthropic are facing an existential threat to their market position.
That assumption misses what is actually happening.
The Real Work of Frontier Models
Decagon CEO Jesse Zhang has proposed a clearer way to understand the split between proprietary and open-source AI. The competition is not a simple race to replace one another. Instead, the two categories serve different phases of a single enterprise lifecycle.
Frontier models function as the discovery layer. When a company begins exploring how AI might change an internal process, the task is almost always poorly defined. The inputs are messy. The desired outputs are vague. Success requires reasoning through ambiguity, handling edge cases the team has not yet catalogued, and adapting to prompts that change by the hour. This is prototyping and proof-of-concept work. It demands the most capable system available, cost be damned, because the alternative is a failed experiment that teaches nothing.
In this phase, an expensive frontier model is not overhead. It is the cost of market research compressed into a few API calls. Once the use case is proven and the workflow is mapped, the nature of the problem shifts. The ambiguity disappears. Inputs become standardized. Prompts stabilize. The task has become routine. At that point, many enterprises migrate the workload to a lighter, cheaper model. Open-source alternatives step in and own the production stage, while frontier models remain stationed at the frontier, handling the next wave of unknown problems.
This is not a theory about how AI should work. It is a description of how budgets are already moving.
When Workloads Move Downmarket
The migration to open source is real, and it is visible in the traffic data. DeepSeek’s surge to over one-third of token volume on Vercel’s infrastructure shows that companies are running enormous quantities of inference through cheaper models. Z.ai’s GLM-5.2 has also carved out a top-four position by handling steady, predictable traffic.
These models excel at tasks that have been tamed. Think of high-volume data extraction from standardized forms, first-pass customer support triage that routes tickets based on obvious keywords, or routine code linting and documentation generation. The prompts are templated. The error modes are understood. The business risk of a bad output is contained. When the work is defined and repetitive, the cost of inference becomes the primary concern. Running that same workload on a six-cent model instead of a premium tier makes immediate financial sense.
But volume is not revenue. The fact that open-source models dominate token counts does not mean they dominate value creation. Token volume measures activity. Token spend measures what companies are willing to pay for irreplaceable capability.
Where the Money Actually Flows
Vercel’s AI gateway data makes the economic split impossible to ignore. Despite DeepSeek’s dominance in raw token traffic, Anthropic continues to capture more than half of the total AI spend on the platform. The gap between activity and expenditure comes down to a staggering price differential.
According to OpenRouter data, Anthropic’s Opus 4.8 costs approximately $1.37 per million tokens. DeepSeek’s V4Flash costs roughly six cents for the same volume. Opus is priced about twenty-three times higher. That multiplier matters more than raw token counts ever could. A development team could move ninety percent of their inference volume to the cheaper model and still see the expensive model account for the majority of their budget.
This discrepancy is not an accident. It reflects the reality that frontier providers are selling something different. They are not just selling tokens. They are selling the ability to reason through problems that lack established playbooks. The enterprises paying premium rates are not doing so out of ignorance. They are doing so because the tasks they assign to these models are either high-stakes or structurally complex. A legal team analyzing novel regulatory exposure cannot tolerate a hallucinated citation. A product team designing a multi-step agentic workflow needs the model to correctly chain logic across several turns. The cost of failure in these scenarios far exceeds the cost of the API call.
Capital expenditure flows toward the layer where the value is still being created, not merely executed.
The Market Grows Faster Than the Migration
If open-source models are so much cheaper, and enterprises are actively shifting mature workloads toward them, why has frontier spending not collapsed? The answer is that the total market of addressable AI tasks is expanding faster than any single model can commoditize it.
Every time a company successfully automates a predictable workflow with an open-source alternative, two things happen. First, that team saves money on execution. Second, those resources and that talent get redirected toward harder adjacent problems. The routine work is now handled by machines, which means the humans can focus on the irregular, the strategic, and the unprecedented. The newly discovered problem almost always requires the reasoning depth of a frontier model.
This pattern repeats across industries. A bank automates document review using a cheap model, then turns its attention to building a dynamic risk model that requires nuanced judgment. A software firm automates test generation, then attempts to build an autonomous debugging agent that must trace errors across distributed systems. The frontier keeps advancing. As soon as one task becomes a commodity, a more complex use case emerges that demands premium capability.
Many enterprise tasks also remain too sensitive to hand to the current generation of open-source alternatives. Medical triage support, financial forecasting under regulatory scrutiny, and executive strategy analysis carry downside risks that make inference cost irrelevant next to accuracy and reliability. These workloads create a durable premium tier. The result is a stable two-tiered economy: a high-margin layer for complex reasoning and discovery, and a high-volume commodity layer for routine production execution.
What This Means for Enterprise Buyers
The practical takeaway is that model selection should follow the maturity of the work, not ideology. Build and validate new AI applications on the most capable frontier models you can access. Pay the premium during discovery. It is cheaper than building on a limited model, failing to prove value, and abandoning the project. Once the inputs, outputs, and failure modes are known, then optimize aggressively. Move the stable workload to an open-source alternative and capture the cost savings.
Trying to force every task into a single tier is a recipe for either wasted capital or missed capability. The companies that navigate this correctly will run hybrid architectures by default, not as a compromise.
The story here is not that open source is losing, or that frontier labs are invincible. It is that both tiers are growing, but they are growing in different directions. Open-source models are swallowing the known world of AI tasks. Frontier models are staking claims on the unknown. For the foreseeable future, that is a comfortable arrangement for both.