Why the Rise of Open Source AI Isn’t Affecting Anthropic … Yet

On Monday, Decagon’s CEO, Jesse Zhang, put forward a compelling theory titled “Everyone is wrong about open source AI in the enterprise.” He delves into an intriguing paradox in the current AI landscape: while mature AI deployments are gravitating towards more streamlined models, even in his own organization, the overall financial commitment to expensive, cutting-edge models remains largely unchanged.

Zhang offers a novel viewpoint on the relationship between frontier and open source models, asserting that they are not competitors. He argues that the rise of open source models does not lessen the significance of frontier labs. Rather, they represent two phases of the same lifecycle, where high-cost frontier models validate use cases that can later shift to more economical open source alternatives as they mature.

As established use cases adopt more streamlined models, new use cases continue to emerge, thereby keeping overall spending on frontier models relatively stable.

Although Zhang provides limited data to back his claims, relevant statistics are readily accessible. Vercel’s AI gateway dashboard reveals that DeepSeek has recently surged to prominence in token volumes, accounting for slightly over one-third of the tokens processed through the company’s infrastructure. Furthermore, Z.ai, the creator of the widely used GLM-5.2 model, has secured a respectable fourth position in the same period.

However, upon reviewing the total expenditure on tokens, Anthropic still commands more than half of the total AI spending on the platform. Much of the recent fluctuations are attributed to Anthropic’s own rising costs; despite a slight drop in share over the past month, the change is not considerable.

OpenRouter presents a comparable scenario, targeting a wider (though slightly less enterprise-centered) segment of the market. DeepSeek V4 Flash emerges as the leading player in terms of usage, processing 5.3 trillion tokens weekly, while the foremost frontier model, Opus 4.8, handles just over 2 trillion tokens. Although OpenRouter does not rank models by total funding, it suggests that the average token cost for Opus 4.8 is roughly 23 times that of V4 Flash ($1.37 per million tokens compared to just 6 cents), indicating that Opus likely continues to garner the majority of expenditures.

These statistics do not account for Nvidia’s latest entry, the Nemotron, which is expected to quickly ascend the ranks due to Nvidia’s extensive connections and the model’s impressive versatility.

While these figures don’t explicitly validate Zhang’s viewpoint on AI life cycles, they do suggest that frontier labs like Anthropic are not currently facing major threats from the rise of open source—at least for the time being. One possible explanation is that the market for AI-suitable tasks is expanding so rapidly that leading models maintain their status by dominating early deployments. As Zhang aptly puts it, “The frontier labs will keep owning discovery. Open source will increasingly own production.” Another potential reason could be that many use cases are complex enough that they cannot be wholly replaced by cheaper alternatives, even as clients shift towards open source.

Regardless, this dual-layer model economy seems set to become a lasting feature of the AI landscape.

Just last September, I discussed the possibility of foundational labs becoming akin to coffee bean suppliers for Starbucks—essentially serving as commodity inputs while the application layer reaps the profits. Some elements of that prediction have come to fruition: Vertical AI companies have transitioned to lighter models, and the economics surrounding “GPT wrapper” startups have mostly remained stable.

Nonetheless, it is clear that, on a token-for-token basis, frontier providers have successfully retained the most coveted segment of the market—the premium token price. This trend does not appear likely to change in the near future.