On Monday, Jesse Zhang, chief executive of Decagon, posted a theory titled “Everyone is wrong about open source AI in the enterprise.” He argues that while his own company is moving to lighter models for mature work, total spending on high-end AI remains steady. This suggests expensive frontier models and cheaper open-source options are not fighting for the same customers but rather serve different stages of a project. Frontier models handle initial discovery, while open-source tools take over once a use case matures and production begins. As new applications emerge, the demand for top-tier models stays high.
The numbers
Zhang offered little data to back his claim, yet the figures are available elsewhere. Vercel’s dashboard shows DeepSeek processing just over a third of all tokens in the past week. Z.ai, the team behind GLM-5.2, moved into fourth place during the same period. Despite these shifts, Anthropic still accounts for more than half of total AI spend on the platform. Their share has dipped slightly due to recent price hikes, but the drop is minimal.
OpenRouter, which covers a broader market segment, shows similar trends. Deepseek V4Flash now processes 5.3 trillion tokens weekly. The leading frontier model, Opus 4.8, handles just over 2 trillion tokens. While OpenRouter does not publish total spend, the average cost for Opus 4.8 is roughly 23 times higher than V4Flash. Opus costs $1.37 per million tokens, compared to just 6 cents for the open-source alternative. This price gap suggests Anthropic continues to capture the majority of the budget.
These stats do not include Nvidia’s Nemotron, which is expected to rise quickly due to the company’s hardware connections and the model’s flexibility.
The data supports Zhang’s view that frontier labs are not suffering significantly from open-source growth, at least for now. The market for AI tasks is expanding fast enough that top models maintain their position by dominating early-stage projects. Zhang noted that frontier labs will keep owning discovery, while open source will increasingly own production. Even as clients switch to cheaper tools, some use cases remain too complex for alternatives to handle fully.
This two-tiered system may become a stable feature of the industry. As recently as September, I wrote about the risk of foundation labs becoming commodity providers, similar to coffee bean suppliers for Starbucks. While some vertical AI players did switch to lighter models, the economics of “GPT wrapper” startups have remained stable. Frontier providers have successfully held onto the premium token price, and there is no sign this will change soon.
What it means
Businesses should expect a permanent split in their tool stacks. You will likely use expensive, high-accuracy models for research and initial prototyping, then migrate to cheaper, open-source versions for running daily operations. The cost of the best models will remain high because the market for early-stage problem solving is growing faster than the cost of the tools themselves.
