Open-source models have won the Token traffic, while Anthropic has taken the majority of the profits.

On Monday, Decagon CEO Jesse Zhang published an article noting that while open-source models are on the rise, their share of total enterprise AI spending is actually declining across the broader corporate market.

This inevitably brings to mind Dario Amodei's previous controversial take: in the AI field, open source is a distraction, or even a false proposition.

"Even when a model is publicly available, you cannot see its internal operational mechanisms, so the industry typically refers to such models as 'open weights' rather than 'open source.' Traditional open-source software can be collaboratively modified, iterated continuously, and improved through community collaboration, but these advantages do not fully apply to large language models," Amodei stated.

As a result, when he encounters a new model, he never first asks whether it is open source. Even whether DeepSeek is open source matters little to him. The only question he truly cares about is: Is this model good enough, and can it outperform us on critical tasks? Moreover, open source never equals free. Models are so large that enterprises still need to spend money running inference on the cloud, and require professional teams for deployment and optimization to ensure the model operates quickly and stably enough. Instead of obsessing over whether a model is open source, it is better to focus on whose model can truly perform the tasks you need completed best.

As AI applications gradually mature, more and more enterprises are shifting to lighter-weight models, and Decagon itself is no exception. However, Jesse notes that overall enterprise spending on expensive frontier models has barely decreased.

According to reports, approximately 90% of Decagon's current workloads run on open-source models, rather than models from OpenAI or Anthropic. He explained:

This is not driven by cost, nor is it something our clients force us to do, although they generally do not object. The real reason is that we have few other viable options.

When you run a customer service AI agent in production, latency directly determines whether the product is usable. No one will use a product where each turn in a conversation takes 8 seconds to receive a response. Therefore, you need smaller, faster models. For each model call, there is no need to know the capital of Lithuania or master high school physics knowledge.

But out-of-the-box small models cannot meet the quality standards clients require. Only through large-scale fine-tuning for specific tasks can they meet those requirements.

The problem is that frontier model labs generally do not offer this combination. We cannot fine-tune their most powerful models according to actual needs, and their small models do not belong to us, so we cannot shape them as we wish.

Therefore, "small model + deep fine-tuning" essentially means you must use an open-weight model. Cost savings do exist, but they are only a secondary benefit; greater enterprise confidence in self-hosted models is also an added bonus, not the fundamental reason we choose open-source models.

In Jesse's view, frontier models and open-source models are not competitors, and the success of open-source models does not come at the expense of frontier model labs losing market share. On the contrary, they are more like two phases in the same lifecycle: enterprises first use expensive frontier models to verify whether an application scenario is viable, and once the scenario matures, migrate it to lower-cost open-source models.

As more mature scenarios shift to lightweight models, new application scenarios are constantly emerging, so overall enterprise spending on frontier models has not decreased significantly.

Jesse Zhang did not provide much data to support this view, but TechCrunch found some relevant statistics.

Vercel's AI Gateway dashboard shows that in just the past week, DeepSeek's token processing volume rapidly rose to first place, now accounting for over one-third of total token traffic on Vercel's infrastructure. Zhipu, the developer of the popular model GLM-5.2, also jumped to fourth place during the same period.

However, when examining overall token spending, Anthropic still accounts for more than half of total AI spending on the platform. Due to a significant portion of recent changes stemming from Anthropic's own price increases, its spending share has slightly decreased over the past month, but the decline is not substantial.

Data from OpenRouter shows a similar trend. Compared to Vercel, OpenRouter covers a larger market, but has a slightly smaller proportion of enterprise users.

In terms of overall usage, DeepSeek V4 Flash is currently the clear leader, processing approximately 5.3 trillion tokens per week. The most popular frontier model, Opus 4.8, processes just over 2 trillion tokens per week.

OpenRouter does not rank models by total spending, but its data shows that the average token cost for Opus 4.8 is approximately 23 times that of V4 Flash: Opus 4.8 averages around $1.37 per million tokens, while V4 Flash costs only $0.06 per million tokens.

Extrapolating from this price gap, even though Opus 4.8's token usage is significantly lower than V4 Flash's, it likely still accounts for the majority of model spending on the platform.

These figures do not yet include the latest entrant, Nvidia Nemotron. Leveraging Nvidia's strong industry relationships and Nemotron's exceptional adaptability, this model is poised to quickly climb the market rankings.

These numbers are not sufficient to fully confirm Jesse Zhang's judgment about the AI application lifecycle, but they do at least indicate that frontier model labs like Anthropic have not been significantly impacted by the rise of open-source models—at least not yet.

One explanation is that the market for tasks that can be automated by AI is growing too rapidly. Even as more mature scenarios are migrated to open-source models, top-tier models can maintain their market position by dominating the early validation phase of new applications.

As Jesse Zhang put it: "The declining share of open-source model spending is not because open-source models are failing, but because the entire enterprise AI market is still in the very early stage of the maturity curve." "Frontier model labs will continue to lead in application discovery, while open-source models will increasingly dominate production deployments."

When an application scenario first emerges, enterprises will prioritize the most capable general-purpose model they can access. Because you do not yet know what form the problem will ultimately take, you are willing to pay a premium for intelligent capabilities that you may not even use in the future. At this stage, this is a reasonable trade-off.

But when an application scenario is fully mature, and enterprises understand the distribution of input data, the behavior the model needs to exhibit, and the failure modes that must be prevented, the trade-off reverses.

At this point, general-purpose intelligence becomes an unnecessary burden. What enterprises truly need is the smallest, fastest model that has been specifically fine-tuned to excel at a particular task.

"Following this logic, every application scenario that currently uses frontier models for prototyping could potentially migrate to open-source models in the future," Jesse stated. "However, this process will take longer than many people expect."

Another explanation is that even as clients begin to shift to open-source models, many application scenarios remain sufficiently complex that they cannot be completely replaced by lower-cost models.

In any case, this dual-tier model economy has the potential to become a relatively stable structure in the AI industry.

References:

https://www.threads.com/@whaleagent/post/DaJx5pfkzmL

https://x.com/thejessezhang/status/2074154325933424861

https://techcrunch.com/2026/07/07/why-the-rise-of-open-source-ai-isnt-hurting-anthropic-yet/

This article is from the WeChat public account "AI Frontline", author: Chu Xingjuan, published with authorization by 36Kr.