Why the rise of open source AI isn’t hurting Anthropic … yet

A new argument from Decagon CEO Jesse Zhang is reframing how people think about open-source AI in the enterprise. In a post published Monday, he challenged the common assumption that cheaper open models are directly undermining frontier AI providers. His point was simpler and more nuanced: companies often begin with powerful, expensive models to validate a use case, then later move mature workloads to lower-cost alternatives. Even so, spending on premium models has remained surprisingly steady.

That creates an apparent contradiction in the AI market. On one hand, organizations with more established deployments are increasingly relying on smaller or open models for production work. On the other, the biggest and most advanced models are still capturing a large share of spending. Zhang’s explanation is that these categories are serving different moments in the same adoption cycle rather than fighting for exactly the same role.

Two different jobs in one AI adoption cycle

Under this view, frontier models and open-source models are less like substitutes and more like sequential tools. High-end models are used first, when teams are still experimenting, testing quality, and figuring out whether a workflow is worth building around. Once those applications are stable and better understood, some of that work can be shifted to lighter and cheaper models.

If that pattern holds, it helps explain why premium model spending has not dropped as sharply as some expected. Older use cases may migrate downward on the cost curve, but new use cases continue to appear. That constant flow of fresh experimentation keeps demand for the top tier alive.

Usage data points in that direction

Zhang did not present extensive data in his post, but platform-level metrics from elsewhere suggest a similar dynamic. On Vercel’s AI gateway dashboard, DeepSeek recently rose to the top in token volume, handling a little more than one-third of the tokens moving through the company’s infrastructure over the past week. During that same period, Z.ai, the team behind GLM-5.2, climbed into fourth place.

Yet token volume tells only part of the story. By spending, Anthropic still appears to dominate on Vercel’s platform, accounting for more than half of AI-related spend. That share has edged down somewhat over the last month, with Anthropic’s own higher pricing contributing to the shift, but the change has been limited rather than dramatic.

OpenRouter shows a related pattern across a broader market, though one that is somewhat less centered on enterprise buyers. There, DeepSeek V4 Flash leads in usage, processing 5.3 trillion tokens per week. The most widely used frontier model on the platform, Opus 4.8, handles a little more than 2 trillion.

OpenRouter does not provide a direct ranking by total spend, but its listed pricing offers a strong clue. Opus 4.8 averages about $1.37 per million tokens, while V4 Flash sits around $0.06 per million. That makes Opus roughly 23 times more expensive per token, which strongly suggests that frontier models may still command most of the revenue even when they trail in raw usage.

Those comparisons also leave out Nvidia’s Nemotron, a model family that could quickly become a major factor thanks to Nvidia’s ecosystem reach and the flexibility of the offering.

Why open source has not seriously weakened frontier labs

None of these numbers fully prove Zhang’s theory, but they do support the broader idea that open-source momentum has not yet translated into major damage for companies at the top end of the market, including Anthropic.

One reason may be the speed at which AI-appropriate work is expanding. If the pool of possible use cases is growing fast enough, frontier labs can continue to thrive by owning the earliest phase of deployment. In Zhang’s framing, “The frontier labs will keep owning discovery. Open source will increasingly own production.”

Another possibility is that many business tasks are still too demanding for lower-cost models to handle on their own. Even if companies adopt open models for some parts of a workflow, they may still depend on frontier systems for harder, less predictable, or more quality-sensitive tasks.

Taken together, that points toward a market with two durable layers rather than a single winner-take-all outcome. The most advanced models may remain the preferred option for exploration and difficult workloads, while cheaper open models absorb more routine production traffic.

Premium pricing remains the key advantage

There was a recent argument that foundation model providers might eventually be reduced to commodity suppliers, with most of the value shifting upward to the application layer. Parts of that thesis have indeed played out. Vertical AI companies have moved some workloads to lighter models, and the economics of many so-called “wrapper” businesses have remained more resilient than critics expected.

But the most important piece of the market has not slipped away from the frontier labs. They still appear able to charge premium rates for premium capability. On a per-token basis, that remains the most attractive section of the business, and there is little evidence so far that it is disappearing anytime soon.