Open source AI is becoming harder to ignore in enterprise software. Lighter models are taking on more production work, and usage dashboards show open alternatives moving enormous token volumes.
Yet that does not mean frontier labs are being pushed out of the market. The emerging picture is more complicated: open source AI may be absorbing mature workloads while expensive frontier models keep winning the early, difficult, and premium parts of the market.
A different way to read the model market
Decagon CEO Jesse Zhang framed the debate with a post titled “Everyone is wrong about open source AI in the enterprise.” His argument addresses a tension at the center of the current AI economy.
According to Zhang, more mature AI deployments are moving to lighter models, including at Decagon. At the same time, overall spending on expensive state-of-the-art models has barely moved. That looks contradictory only if open source and frontier models are treated as direct substitutes.
Zhang’s view is that they may instead be different stages of one life cycle. Frontier models help teams test and validate new AI use cases. Once those use cases become better understood, some can move to cheaper open source alternatives.
That shift does not automatically reduce spending on frontier models. As older use cases move down to lighter systems, new use cases keep appearing. The frontier model budget can remain meaningful because the most demanding work keeps moving to the edge of what current models can do.
Usage is shifting faster than spending
The available dashboard data described in the source supports the basic shape of this argument, even if it does not fully prove the full life-cycle theory.
Vercel’s AI gateway dashboard shows DeepSeek leading token volume over the past week, processing just over a third of the tokens moving through that infrastructure. Z.ai, the lab behind the popular GLM-5.2 model, reached fourth place over the same period.
But token volume is not the same as token spend. On Vercel’s dashboard, Anthropic still accounts for more than half of overall AI spend on the platform. Its share has dipped slightly over the past month, with much of the recent change tied to Anthropic’s own rising prices, but the shift is not described as significant.
OpenRouter shows a similar split between usage and economics. DeepSeek V4 Flash leads overall usage there, processing 5.3 trillion tokens weekly. The most popular frontier model, Opus 4.8, handles just over 2 trillion.
OpenRouter does not rank models by total spend, but it lists the average token cost for Opus 4.8 as roughly 23x higher than V4 Flash. The source gives those prices as $1.37 per million tokens for Opus 4.8, compared to just 6 cents for V4 Flash. On that basis, Opus was still probably taking the largest share of spending.
Why frontier labs still have leverage
The split suggests a two-tiered AI market. One tier optimizes for cost, scale, and production efficiency. The other tier pays for the highest capability available when teams are still exploring what a model can do.
That structure helps explain why the rise of open source AI is not necessarily hurting Anthropic yet. Open models can gain huge token volume while frontier providers keep the premium token price.
Several forces described in the source point in the same direction:
- Mature deployments can become cheaper. Once a task is understood, lighter models may be good enough to handle it.
- New use cases keep appearing. If the market of AI-addressable tasks grows quickly, frontier models can keep finding fresh demand.
- Some work remains hard to replace. Even when clients adopt open source, difficult use cases may still need more expensive models.
- Price matters as much as volume. A model with fewer tokens can still capture more money if each token costs far more.
Zhang summarized the division directly: “The frontier labs will keep owning discovery. Open source will increasingly own production.”
That line captures the core economic distinction. Discovery is where teams try to prove whether a new AI workflow works at all. Production is where a known workflow needs to run repeatedly, reliably, and at lower cost.
Nvidia’s Nemotron adds pressure to the open side
The data points in the source do not include the newest arrival, Nvidia’s Nemotron. Still, the article notes that Nemotron is positioned to move quickly because of Nvidia’s strong connections and the model’s extreme adaptability.
That matters because open source AI is not a static category. If new models keep arriving with stronger distribution and broader flexibility, the production side of the market can keep expanding. More workloads that once required a frontier model may become candidates for lighter alternatives.
But that still does not guarantee a collapse in frontier spending. If every wave of cheaper models also encourages companies to find new AI tasks, then the most capable systems can continue to serve the newest and hardest work.
The premium layer has not disappeared
The source frames this as a shift from an earlier concern about foundation labs becoming commodity suppliers. As recently as last September, the author was writing about the possibility that foundation labs would end up selling basic inputs while the application layer captured more of the value.
Some of that prediction appears to have happened. Vertical AI companies have moved to lighter models, and the economics of “GPT wrapper” startups have stayed mostly stable.
But the other side of the prediction has not fully arrived. Token for token, frontier providers have kept the most desirable part of the market: the premium token price. That is the central reason open source AI can grow quickly without immediately damaging Anthropic’s position.
The result is not a simple victory for one side. Open source models are becoming more important for production. Frontier labs are still important for exploration, difficult tasks, and high-priced usage.
For now, the AI economy appears to be sorting itself into layers rather than replacing one model category with another. Open source AI may be winning more of the workload, but Anthropic and other frontier labs still appear to be holding the part of the market where each token is worth the most.