Open source AI models are gaining ground fast in usage, but new evidence suggests the boom is not yet taking a meaningful bite out of Anthropic’s enterprise revenue. Instead, the market appears to be splitting into two layers: cheaper open source systems for mature production workloads, and expensive frontier models for the hardest, earliest-stage work.
open source ai
A theory of two AI markets
The idea comes from Decagon CEO Jesse Zhang, who on Monday published a post titled “Everyone is wrong about open source AI in the enterprise.” His argument centers on a contradiction in today’s AI economy: companies are increasingly switching some deployments to lighter models, including at his own company, but spending on top-tier models has barely changed.
In Zhang’s view, open source models are not simply replacing frontier labs. Instead, the two categories may represent different phases of the same product lifecycle. Companies begin with expensive, state-of-the-art models to discover what works, then move those established use cases to cheaper alternatives once they are stable enough for production.
That would mean open source success does not necessarily come at the expense of frontier providers. As more mature workflows shift downward, new and more demanding use cases continue to emerge, keeping spending on the most advanced models high.
Usage is shifting, but spend is not collapsing
Zhang’s post does not include much supporting data, but the broader market seems to back up the pattern. Vercel’s AI gateway dashboard shows DeepSeek climbing sharply in token volume over the past week, now handling just over a third of the tokens flowing through Vercel’s infrastructure. Z.ai, the lab behind the GLM-5.2 model, also moved up the rankings and landed in fourth place.
Yet when the dashboard is sorted by overall token spend, Anthropic still accounts for more than half of the platform’s AI spend. That share has slipped slightly over the past month, partly because Anthropic has raised prices, but not enough to suggest a major erosion in demand.
The result is a clear split between raw usage and monetization: open source and lower-cost models may be handling more tokens, but frontier models still appear to command the most valuable workloads.
OpenRouter shows a similar pattern
OpenRouter, which covers a broader and less enterprise-focused slice of the market, tells much the same story. DeepSeek V4 Flash is now the top model by overall usage, processing 5.3 trillion tokens per week. By comparison, the most popular frontier model, Opus 4.8, handles a little more than 2 trillion weekly tokens.
OpenRouter does not rank models by total spend, but its pricing data suggests why frontier labs can still dominate revenue even when they lose share in usage. The average token cost for Opus 4.8 is roughly 23 times higher than DeepSeek V4 Flash: $1.37 per million tokens versus 6 cents per million. That price gap means Opus is still likely capturing the bulk of spending.
There is also a new entrant to watch. Nvidia’s Nemotron is not yet reflected in those figures, but the source material suggests it could quickly rise to the top thanks to Nvidia’s strong distribution relationships and the model’s flexibility.
Why Anthropic still looks insulated
The current numbers do not prove Zhang’s thesis outright, but they do support a more cautious conclusion: frontier labs such as Anthropic are not being badly hurt by the rise of open source models, at least not yet.
One explanation is simple market growth. The universe of AI-addressable tasks is expanding so quickly that frontier providers can maintain their position by winning early deployments, even if some of those customers later migrate to cheaper systems for routine production use.
As Zhang put it, “The frontier labs will keep owning discovery. Open source will increasingly own production.”
Another explanation is that many high-value use cases still remain too complex to be fully replaced by cheaper models. In those cases, customers may adopt open source tools for some tasks, but continue paying premium prices for the most difficult work, where accuracy, flexibility, or reliability matter most.
A durable two-tier model economy
That dynamic suggests the AI market may settle into a stable two-tier structure. Frontier models would remain the expensive layer used to explore new applications and solve difficult problems, while open source models would increasingly power established, high-volume production systems.
This is not a new worry for AI labs. As recently as last September, the possibility that foundation model companies would become commodity suppliers to the application layer was already being discussed. Some of that has happened. Vertical AI startups have shifted to lighter models, and the economics of “GPT wrapper” businesses have remained largely intact.
But the pricing power at the top has not disappeared. Even as cheaper models gain volume, frontier providers still seem able to hold onto the premium token prices that matter most to revenue. For now, that suggests open source AI is changing how models are used more than it is changing who gets paid.
Source: Original report
Was this helpful?
Last Modified: July 8, 2026 at 10:55 am
1 views
