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

On Monday, Decagon CEO Jesse Zhang published a provocative new theory, posted under the title “Everyone is wrong about open source AI in the enterprise.” The article grapples with one of the most interesting contradictions of today’s AI economy: more mature AI implementations are switching to lighter models, he says, even in his own company. But overall spending on expensive, ultra-modern models has barely budged.

It’s a new way to think about the relationship between frontier and open source models. According to Zhang, they are not competitors, and the success of open source models does not come at the expense of frontier laboratories. Instead, they are two phases of the same lifecycle, using expensive frontier models to prove use cases that can be passed on to cheaper open source alternatives as they mature.

As more mature use cases move to lighter models, new use cases continue to emerge – and overall spend on frontier models is barely budging.

Zhang doesn’t provide much data to support this point, but the data isn’t hard to find. Vercel’s AI gateway dashboard shows that DeepSeek has taken the lead in token volumes over the past week and now processes just over a third of the tokens moving through the company’s infrastructure. Z.ai – the lab behind the popular GLM-5.2 model – jumped to a respectable fourth place in the same period.

But if you scroll down to total token spend, you’ll see that Anthropic is still responsible for more than half of the total AI spend on the platform. Since much of the recent change comes from Anthropic’s own rising prices, the stock is down slightly over the past month, but not significantly.

OpenRouter tells a similar story and covers a much larger (but slightly less entrepreneurial) segment of the market. DeepSeek V4 Flash is the top winner in terms of total usage, processing 5.3 trillion tokens every week. The most popular frontier model, Opus 4.8, handles just over 2 trillion. OpenRouter doesn’t rank models by total spend, but records the average token cost for Opus 4.8 as about 23x higher than V4 Flash ($1.37 per million tokens, compared to just 6 cents), which would mean that Opus likely still accounted for the lion’s share of the spend.

These numbers don’t even include the newest addition, Nvidia’s Nemotron ready to jump to the front of the pack thanks to Nvidia’s strong connections and the model’s extreme adaptability.

These numbers don’t fully prove Zhang’s point about AI lifecycles, but they do show that frontier labs like Anthropic aren’t suffering too much from the rise of open source – at least not yet. One explanation is that the market for AI-addressable tasks is growing so quickly that the top models can maintain their position by dominating early-stage deployment. As Zhang puts it, “The frontier labs will continue to own discoveries. Open source will increasingly own production.” Another explanation could be that even as customers move to open source, many use cases are so difficult that they cannot be completely replaced by cheaper alternatives.

Be that as it may, this dual economy of models may become a relatively stable feature of the AI ​​economy.

Last September, I wrote about the possibility that foundation labs would eventually sell coffee beans to Starbucks—that is, they would serve as a raw material while the application layer reaped the benefits. Some parts of that prediction have come true: vertical AI games have moved to lighter models, for example, and the economics of “GPT wrapper” startups have remained largely stable.

But we also see that, token by token, frontier providers have managed to hold on to the most desirable part of the market: the premium token price. And it doesn’t seem likely that this will change anytime soon.