Open Source AI Is Surging, But Anthropic Still Owns the Premium Tier

Open-source AI models are gaining ground fast in enterprise and developer workflows, but the rise of cheaper alternatives has not yet knocked Anthropic off its perch. New usage data from model-routing platforms suggests the market is splitting into two distinct layers: open models are increasingly taking over high-volume production workloads, while frontier models continue to command the most expensive, strategically important early-stage work.

That dynamic is reshaping one of the biggest assumptions in artificial intelligence: that every jump in open-source capability should automatically eat into the business of the leading model labs. In practice, the picture looks more complicated. Enterprise buyers are optimizing for cost on some tasks, but they are also adding new AI use cases quickly enough that overall spending on premium models remains stubbornly high.

The new AI split: discovery versus production

The latest argument comes from Decagon chief executive Jesse Zhang, who described open source and frontier models not as direct rivals, but as successive stages in a product life cycle. In his view, premium models are often used first to validate a use case, define the workflow and prove value. Once that happens, the workload can migrate to a cheaper open model for scale and efficiency.

That framing helps explain a strange contradiction in the market: more mature deployments are indeed moving to lighter models, yet the total amount of money flowing into frontier systems has not fallen dramatically. Instead of replacing premium models outright, open source appears to be taking over one slice of the market while the frontier segment keeps finding new demand.

“The frontier labs will keep owning discovery. Open source will increasingly own production,” Zhang argued in his post on enterprise AI deployment.

His point is less about ideology than economics. The highest-value use cases in AI are still expensive to run, harder to replace, and more likely to require advanced reasoning, reliability and vendor support. Open-source systems may be winning usage volume, but that does not necessarily mean they are winning the most valuable work.

What the usage data shows

Evidence from AI infrastructure providers supports that reading. Vercel’s AI gateway dashboard, which tracks traffic moving through its systems, shows that DeepSeek has recently surged to the top of the platform by token volume, handling just over one-third of the tokens processed in the past week. Z.ai, the company behind the GLM-5.2 model, has also climbed quickly and now sits near the top tier of active usage.

But token volume is only part of the story. When the same dashboard is sorted by total spend, Anthropic still accounts for more than half of AI-related costs on the platform. That is a significant gap. It suggests that while open-source models may be processing more requests, the more expensive frontier systems are still monetizing a larger share of the value flowing through enterprise applications.

The balance has shifted slightly over the past month, in part because Anthropic has raised prices. Even so, the company’s dominance on spend has not materially eroded. In other words, the market may be diversifying in usage, but premium demand remains resilient.

Why tokens and dollars tell different stories

Token volume measures activity. Spend measures revenue capture. The two are related, but they can point in very different directions.

• A cheap model can process vast volumes of low-margin work and still generate less revenue.
• An expensive model can handle fewer tokens and still collect far more money.
• Enterprise buyers may use frontier models for planning, debugging, evaluation or other high-stakes tasks, then shift repetitive workloads to lower-cost models.

This split helps explain why open source can “win” usage without yet devastating frontier labs. A model that is used more often is not automatically the model that earns more.

OpenRouter points to the same pattern at larger scale

A similar trend appears in OpenRouter, a routing platform that serves a broader mix of developers and companies. There, DeepSeek V4 Flash has emerged as the major usage leader, processing about 5.3 trillion tokens each week. That is an enormous amount of activity, and it underscores how quickly efficient open models can become the default option for many routine tasks.

Anthropic’s Opus 4.8, meanwhile, remains the most popular frontier model on the platform and processes a little more than 2 trillion tokens weekly. On raw usage, it trails the leading open model by a wide margin. Yet OpenRouter’s pricing data tells another story: Opus 4.8 is listed at roughly $1.37 per million tokens, compared with 6 cents for DeepSeek V4 Flash — a difference of around 23 times.

That pricing gap means Anthropic is still positioned to capture a disproportionate share of spending even when it processes fewer tokens. The exact revenue split is not publicly visible on OpenRouter, but the economics are clear enough: the premium model can keep monetizing a smaller slice of demand far more effectively than a low-cost alternative.

Why frontier labs are not feeling the pain yet

If open source is moving so quickly, why has it not squeezed out the most expensive vendors? The simplest answer is that the total market is expanding fast enough to absorb both layers at once. Every time a company adopts AI for a new workflow, it creates demand for experimentation, evaluation and integration. Those early stages often favor top-tier models, which are better at handling ambiguous prompts, complex instructions and fragile enterprise requirements.

At the same time, some workloads are still too difficult to fully hand off to cheaper systems. Enterprises may trust a frontier model for planning, document analysis, customer interactions or other mission-critical tasks even after open-source alternatives become available. Reliability, latency, safety controls and support often matter as much as raw performance.

There is also a strategic reason that premium providers can remain strong: they usually own the earliest and most ambitious deployments. If frontier models help customers discover what is possible, they can remain embedded in the center of the workflow even as production volumes drift elsewhere.

The growth of AI demand changes the math

AI spending is not a fixed pie. As businesses build more internal tools, customer-facing assistants and agentic workflows, they are not merely swapping one model for another. They are adding new use cases altogether. That means model substitution can happen at the same time as market expansion.

In that environment, a lower-cost model can win on throughput without reducing the total revenue opportunity for a premium rival. The market can become layered rather than zero-sum.

A recurring prediction about AI commoditization

The current debate echoes an earlier argument in the industry: that foundation model companies could end up looking more like commodity suppliers than the owners of the biggest profits. A year ago, that concern was often framed in blunt terms, with model makers compared to manufacturers of generic ingredients while application companies captured the customer relationship and the margins.

Some of that thesis has played out. Vertical AI companies have increasingly moved to lighter models once their products are stable, and many startups built on top of large language models have found that their economics are more durable than skeptics expected. But the premium layer has not disappeared. Instead, the market seems to have formed a hierarchy.

In the market’s current shape, frontier models often help establish the use case, while cheaper models take over repeatable production work.

That arrangement is good news for buyers looking to reduce costs. It is not yet bad news for the leading model labs, especially those that can charge premium rates for the most demanding workloads.

What this means for Anthropic

Anthropic is one of the clearest examples of a company benefiting from the premium tier of the AI market. Even as lower-cost models gain traction, the company continues to anchor the expensive end of major routing platforms. That matters because spend, not just usage, is what turns model demand into a business.

Anthropic’s position also highlights a subtler point: in AI, technical leadership and commercial capture do not always move together. A model may lose share in token volume and still remain the main revenue engine in the ecosystem. The gap between those metrics can stay wide for a long time, especially when enterprise customers treat frontier systems as the standard for their most sensitive tasks.

Three reasons Anthropic remains protected for now

1. High-value early deployments: Frontier models are still the preferred choice for testing and launching new AI workflows.
2. Premium pricing: Even small usage shares can translate into significant spend when the per-token cost is much higher.
3. Task complexity: Some jobs are not easily transferred to cheaper models without losing quality or reliability.

That does not mean Anthropic is insulated forever. If open models keep improving, more workloads will migrate. But the current data suggests the transition is likely to be gradual rather than abrupt.

The arrival of Nvidia’s Nemotron adds another wrinkle

The market may also be approaching a new competitive phase with the emergence of Nvidia’s Nemotron. While it is not yet fully reflected in the usage figures cited by routing platforms, the model is widely expected to climb quickly. Nvidia has obvious distribution advantages, deep enterprise relationships and the ability to influence how its models are adopted across hardware and software ecosystems.

Nemotron is notable not just because of the company behind it, but because of its flexibility. Models that can be adapted to a wide range of settings often become attractive defaults in production environments, especially if they offer a strong balance between quality and cost. If that happens, the open-source ecosystem may gain another powerful argument for scale.

Still, the broader lesson may remain unchanged: gaining usage is not the same as displacing spending. A model can become the default production layer while frontier providers continue to dominate discovery and premium workflows.

How the enterprise AI market may settle

The most likely near-term outcome is a two-tier AI economy. In that system, frontier labs would sell the tools companies rely on when the stakes are highest, while open-source models would handle the repetitive, cost-sensitive work that emerges after workflows are proven.

That structure could prove stable because it aligns with how businesses adopt technology. Companies typically start with a more capable, more expensive option while they figure out what actually works. After the workflow matures, they optimize for cost. By then, however, they may already be using AI in more places, not fewer.

Potential outcomes for model providers

• Frontier labs: retain premium pricing and early-stage influence.
• Open-source labs: win volume, production deployment and developer mindshare.
• Enterprises: gain more bargaining power and lower inference costs over time.
• Infrastructure platforms: become the main battleground for routing, pricing and model selection.

That would make the AI economy feel less like a single market and more like a stack of interlocking markets, each with its own economics. Discovery, evaluation, production and optimization may increasingly belong to different kinds of models.

The bottom line

For now, the rise of open-source AI is not inflicting major damage on Anthropic or the other frontier labs. Usage data shows open models are gaining speed, and in some cases dominating token volumes. But spending data shows the expensive systems still control the richest part of the market.

That could change if open models keep improving while enterprises become more aggressive about cost control. It could also change if one of the major frontier labs loses its technical edge. But at the moment, the evidence points to coexistence, not collapse.

Open source is increasingly the engine of production. Frontier models still own the expensive first pass. And as long as AI adoption keeps broadening, there may be room for both.