Chinese Open-Source Models vs OpenAI & Anthropic: Cost, Performance, and U.S. Adoption Benchmarks

The Token Share Numbers Tell the Story

On OpenRouter, a platform that allows developers to access a range of AI models through a single interface, the share of tokens consumed by U.S. companies on Chinese AI models has sat above 30% every week since early February. That figure has climbed as high as 46%. The contrast with prior periods is stark: the 12-month average before this shift was 11%, and in the first half of 2025 it sat at just 4.5%.

That is not a gradual drift. That is a structural reallocation of workloads.

The same pattern appears on Vercel, a platform for deploying and running applications. DeepSeek’s share of gateway tokens climbed between May and June. Z.ai’s GLM 5.2, released in June, recorded the fastest adoption of any model tracked by Vercel in 2026. In its first full week after launch, daily token volume grew approximately 27x and the number of customers using it grew approximately 80x.

These are not vanity metrics from a press release. They reflect engineers making active routing decisions in production environments.

What Is Driving the Switch

The primary driver is cost. Open-source Chinese models can be 60% to 90% cheaper than leading Anthropic and OpenAI models, according to data from OpenRouter. For companies running AI at scale—processing millions of tokens daily across internal tools, customer-facing products, and agentic workflows—that differential is not marginal. It is the difference between a viable unit economics model and one that does not close.

Kyle Chan, a fellow at the Brookings Institution’s John L. Thornton China Center, framed the shift precisely: where U.S. companies previously prioritized AI adoption regardless of which model they used, they are now becoming cost-conscious. The experimentation phase is over. The optimization phase has begun.

AI startup Lindy made this calculation explicit. In June, the company moved 100% of its traffic from Anthropic’s Claude models to DeepSeek. CEO Flo Crivello described the cost curve as crashing to the ground, with the decision projected to save the company millions of dollars within months. Switching to DeepSeek V4 also increased performance on many of Lindy’s core use cases—cost reduction and capability improvement occurring simultaneously.

That combination is what makes this moment different from previous open-source cycles, where cheaper typically meant meaningfully worse.

The Performance Gap Is Narrowing

Chinese models are not yet at parity with the top U.S. frontier systems on every dimension. Chan estimates they are currently six to nine months behind the leading American models. But that framing obscures something important: for the majority of enterprise workloads, six to nine months behind the frontier is more than sufficient.

GLM 5.2 landed within a percentage point of Anthropic’s Opus 4.8 on one closely watched agentic benchmark—at roughly a fifth of the cost. Some researchers have noted that GLM 5.2 performs on par with top U.S. labs on certain cyber benchmarks. These are not claims from the model’s developers; they are observations from practitioners and researchers evaluating outputs.

Justin Summerville, who works on data and analytics at OpenRouter, summarized the practical reality: the new open-source models are performing well and prove capable for all but the most complex LLM tasks.

That qualifier matters. There remains a category of high-complexity, high-stakes tasks where the gap between Chinese open-source models and U.S. frontier systems is meaningful. Legal reasoning at the edge of ambiguity, highly nuanced long-context synthesis, and certain specialized professional domains may still favor the most capable proprietary models. But the proportion of enterprise workloads that genuinely require that level of capability is smaller than many AI vendors would prefer their customers to believe.

How Enterprises Are Actually Deploying This

The emerging pattern is not wholesale replacement of OpenAI or Anthropic. It is intelligent routing—directing tasks to the cheapest model that is good enough for the job.

Harpreet Arora, head of agentic infrastructure at Vercel, described it directly: when a task does not need the best model, teams are beginning to route it to the cheapest one that is good enough, and the recent wave of Chinese models is winning that trade.

On LaunchLemonade, an AI agent platform serving regulated industries, Claude and ChatGPT still dominate overall usage. But GLM 5.2 has entered the top five models on the platform. CEO Cien Solon noted that businesses with more mature AI strategies are increasingly willing to use Chinese models where they make technical or commercial sense—particularly for specific workloads where the performance-to-cost ratio is favorable.

This is the behavior of organizations that have moved past the proof-of-concept stage. They are not asking whether AI works. They are asking which model, at which price point, for which task.

The Open-Source Distinction Matters

It is worth being precise about what “Open-source” and “open-weight” mean in this context, because the distinction carries real implications for enterprise buyers.

Open-source and open-weight models make different components of the model available for inspection, use, and sometimes modification. This is structurally different from closed proprietary systems like many flagship models from OpenAI, Anthropic, and Google, where the code and inner workings remain inaccessible.

For enterprises, this distinction translates into control. Companies can self-host open-weight models, fine-tune them on proprietary data, and avoid dependency on a vendor’s pricing decisions or API availability. Yacine Jernite, head of machine learning at Hugging Face, identified this as a core motivator: companies are increasingly drawn to cheaper AI stacks they can control and adapt themselves, and given the current state of open-source and open-weight models, that often means leveraging Chinese options.

Jernite also articulated the structural risk embedded in the current landscape: users may find themselves forced to choose between performant but expensive U.S. proprietary models—whose price and accessibility can fluctuate—or Chinese models as the only feasible alternative when they want to control costs or own their AI stack. That is not a comfortable binary for enterprises with compliance obligations or geopolitical sensitivities.

The Regulatory Dimension

The rise of Chinese open-source models is not occurring in a policy vacuum. The U.S. administration has been actively examining how to regulate its most powerful AI models and how to slow the adoption of overseas alternatives.

In late June, OpenAI limited the rollout of a new set of models at the government’s request. Export controls on Anthropic’s Mythos and Fable models were lifted that same month, following a standoff between the Trump administration and the company. These are not background events. They are signals that the regulatory environment around AI model access is actively contested and subject to rapid change.

For enterprise buyers, this introduces a new variable into model selection decisions. Proprietary U.S. models are subject to government-directed access restrictions. Chinese models carry their own set of geopolitical and data governance considerations. Neither option is without risk, and the risk profile of each is evolving.

What This Means for Tool Selection

For teams currently evaluating their AI stack, several practical implications follow from this trend.

The cost differential between Chinese open-source models and U.S. proprietary models is large enough to warrant serious evaluation, not just for cost savings but for the operational flexibility that self-hosted or open-weight models provide. The performance gap, while real, is narrowing and is already irrelevant for a significant portion of standard enterprise workloads.

Routing logic—the ability to direct different task types to different models based on cost and capability thresholds—is becoming a meaningful infrastructure decision rather than a nice-to-have. Platforms that support multi-model routing are positioned to benefit from this shift.

The binary framing of “U.S. models vs. Chinese models” is less useful than a workload-level analysis. The more precise question is: which tasks in your workflow genuinely require frontier-level capability, and which can be handled adequately by a model that costs a fraction of the price?

Companies that answer that question rigorously will find that the proportion of workloads requiring the most expensive models is smaller than their current spend implies.