Why the rise of open source AI isn’t hurting Anthropic’s bottom line — yet

As open-source AI models like DeepSeek and GLM-5.2 surge in usage, a natural question arises: Are they eating into the revenue of premium frontier labs such as Anthropic? A new analysis by Decagon CEO Jesse Zhang, combined with fresh data from platforms like Vercel and OpenRouter, suggests the answer is more nuanced than a simple yes or no.

The lifecycle theory of AI model adoption

Zhang’s central argument, published earlier this week, posits that frontier and open-source models are not direct competitors but rather serve distinct phases of a single adoption lifecycle. In this model, expensive frontier models are used to validate and prove out new use cases. Once a workflow is mature and well-understood, organizations often migrate it to cheaper, lighter open-source alternatives to reduce costs.

This pattern explains a seemingly contradictory trend: overall token volume for open-source models is climbing rapidly, yet total spending on frontier models remains resilient. Zhang describes it as a two-tier economy where frontier labs “own discovery” and open-source models “own production.”

What the data shows: usage vs. spend

Data from Vercel’s AI gateway dashboard illustrates the divergence. Over the past week, DeepSeek surged to process just over a third of all tokens on the platform, with Z.ai’s GLM-5.2 jumping to fourth place. However, when measuring actual dollar spend, Anthropic still accounts for more than half of total AI expenditure on Vercel, even after accounting for recent price increases that slightly reduced its share.

OpenRouter, which tracks a broader segment of the market, tells a similar story. DeepSeek V4Flash leads in raw token volume at 5.3 trillion weekly tokens, compared to Opus 4.8’s 2 trillion. But Opus 4.8 costs roughly 23 times more per million tokens — $1.37 versus $0.06 — meaning the frontier model still captures the majority of spending.

Why frontier labs aren’t feeling the pressure yet

Several factors explain this resilience. First, the overall market for AI-addressable tasks is expanding rapidly. New use cases emerge continuously, and frontier models are the default choice for early-stage experimentation. As Zhang puts it, “The frontier labs will keep owning discovery. Open source will increasingly own production.”

Second, many enterprise use cases are sufficiently complex that they cannot be entirely replaced by cheaper alternatives. High-stakes applications requiring reliability, safety, and advanced reasoning continue to rely on premium models, sustaining demand at the high end of the market.

Implications for the AI economy

This two-tiered structure may become a stable feature of the AI landscape rather than a temporary transition. Earlier predictions that foundation labs would become commodity suppliers, akin to selling coffee beans to Starbucks, have only partially materialized. Vertical AI applications have indeed moved to lighter models, but frontier providers have retained pricing power for premium tokens.

The arrival of new models like Nvidia’s Nemotron, which combines strong performance with deep industry connections, could further intensify competition at both tiers. However, the fundamental dynamic — frontier models for discovery, open-source for production — appears durable for the foreseeable future.

Conclusion

While open-source AI models are winning the battle for token volume, frontier labs like Anthropic are not yet losing the war for revenue. The market is large enough and growing fast enough to support both tiers, with each serving a distinct role in the AI adoption lifecycle. For now, the premium token price remains the most valuable real estate in the AI economy.

FAQs

Q1: Are open-source AI models cheaper than frontier models?
Yes. Open-source models like DeepSeek V4Flash cost roughly $0.06 per million tokens, while frontier models like Anthropic’s Opus 4.8 cost around $1.37 per million tokens — a difference of about 23x.

Q2: Why don’t companies just use open-source models for everything?
Many use cases require the reliability, safety, and advanced reasoning capabilities of frontier models, especially during early-stage experimentation and for high-stakes applications where errors are costly.

Q3: Will this two-tier model last?
Current data and expert analysis suggest it may become a stable feature of the AI economy, with frontier models dominating discovery and open-source models handling mature production workloads.