DeepSeek is leading a sharp increase in the use of low-cost Chinese open-weight artificial intelligence models on a major U.S. web development platform, highlighting a rapid shift in developer preferences toward cheaper AI systems that can be deployed with greater flexibility. Data from Vercel’s AI Gateway showed open-weight models accounting for 54% of token volume on Tuesday, August 25, surpassing proprietary models at 46%.
The change represents a dramatic reversal from June, when open-weight models accounted for just 28% of token volume on the same platform, compared with 72% for closed models. Open models subsequently reached a record 62% share on Saturday, according to Vercel CEO Guillermo Rauch. DeepSeek-V4-Flash was the most-used model by token volume on Tuesday, while Chinese models occupied four of the top five positions.
Open-Weight Models Take Majority Share On Vercel
Vercel’s AI Gateway data provides an indication of how developers are choosing models for real-world workloads. Rather than measuring downloads or registrations, the platform tracks token consumption, offering a view into the models actually being used through its infrastructure.
On August 25, open-weight models represented 54% of total token volume, while proprietary models represented 46%. The result was even more pronounced on August 22, when open models reached 62%.
Vercel AI Gateway Model Share
| Date | Open-Weight Models | Proprietary Models |
|---|---|---|
| June 24, 2026 | 28% | 72% |
| August 22, 2026 | 62% | 38% |
| August 25, 2026 | 54% | 46% |
The shift is significant because it occurred within roughly two months. It suggests developers are increasingly willing to use models whose weights are available for download and deployment rather than relying exclusively on closed systems operated by companies such as Anthropic, OpenAI and Google.
DeepSeek-V4-Flash Emerges As The Most-Used Model
DeepSeek was at the center of the shift.
According to Vercel data cited by the South China Morning Post, DeepSeek-V4-Flash was the most-used model on the platform by token volume on Tuesday. Its updated 0731 version ranked fifth, while other Chinese models also occupied leading positions.
The top-five ranking showed the strength of Chinese model providers:
| Rank | Model | Developer |
|---|---|---|
| 1 | DeepSeek-V4-Flash | DeepSeek |
| 2 | DeepSeek-V4-Flash 0731 | DeepSeek |
| 3 | GPT-5.6 Luna | OpenAI |
| 4 | Step 3.7 Flash | StepFun |
| 5 | GLM-5.2 | Zhipu / Z.ai |
The concentration is notable because it means the shift is not being driven by a single Chinese model. Several Chinese AI developers are gaining usage simultaneously.
Why Lower Cost Is Driving Adoption
One of the most important factors behind the change is cost.
Developers building autonomous AI agents can generate extremely large numbers of tokens because agents may repeatedly reason, write code, call tools, inspect results and revise their outputs.
That makes the cost of each token increasingly important.
A model that is slightly less capable but substantially cheaper can become more attractive for high-volume workloads, particularly when developers can use it for routine tasks while reserving expensive frontier models for more complex problems.
Cost Becomes More Important For Agentic AI
AI Agent
↓
Reasoning
↓
Tool Call
↓
Result Analysis
↓
New Reasoning
↓
Additional Tool Call
↓
More Tokens
Unlike a traditional chatbot interaction, an autonomous agent may generate many rounds of model activity before completing a task. As token usage increases, even relatively small differences in per-token pricing can translate into significant operating-cost differences.
The SCMP report specifically links the rise of open-weight models to developers choosing cheaper systems for production workloads, particularly autonomous agents that consume large numbers of tokens.
Chinese Models Gain Ground Beyond DeepSeek
DeepSeek is not alone in benefiting from this trend.
StepFun’s Step 3.7 Flash and Zhipu’s GLM-5.2 were also among the five most-used models by token volume on Vercel’s platform. That suggests the shift toward Chinese AI is becoming broader rather than being entirely dependent on DeepSeek.
Chinese developers have increasingly emphasized efficient architectures, lower inference costs and open-weight releases. These characteristics can make models attractive to developers that want more control over deployment and economics.
Chinese AI Models In The Vercel Top Five
| Company | Model | Position |
|---|---|---|
| DeepSeek | V4-Flash | 1 |
| DeepSeek | V4-Flash 0731 | 5 |
| StepFun | Step 3.7 Flash | 4 |
| Zhipu / Z.ai | GLM-5.2 | 5* |
*The ranking in the source places GLM-5.2 fourth and DeepSeek-V4-Flash 0731 fifth; the table should therefore be read as model presence rather than a separate ranking sequence.
The broader pattern is clear: Chinese open-weight models are increasingly competing for actual developer workloads on infrastructure operated in the United States.
Open-Weight Versus Proprietary AI
The surge represents a broader change in the AI industry’s competitive structure.
Proprietary models are controlled by their developers. Users generally access them through APIs or consumer applications, while the underlying model weights remain unavailable.
Open-weight models allow developers to obtain the model parameters and, depending on licensing terms, run or modify the systems themselves.
Key Differences
| Factor | Open-Weight Models | Proprietary Models |
|---|---|---|
| Model weights | Available | Not publicly available |
| Deployment | Can offer self-hosting | Usually provider-hosted |
| Customization | Greater flexibility | Provider-controlled |
| Infrastructure control | Higher | Lower |
| Upfront complexity | Potentially higher | Lower |
| Cost optimization | More deployment options | Dependent on provider pricing |
| Data control | Can support private deployment | Usually handled through provider |
Open weights do not automatically mean that a model is free, however. Organizations still need computing resources, engineering expertise and infrastructure to run models themselves.
For developers using hosted versions through platforms such as Vercel, the practical benefit can be access to lower-cost model options without having to build the entire serving infrastructure themselves.
The June Reversal Shows How Quickly AI Preferences Can Change
The speed of the shift is perhaps more important than the exact percentage.
On June 24, open-weight models represented only 28% of token volume on Vercel’s AI Gateway. Closed models dominated with 72%.
By August 22, the relationship had completely reversed, with open models reaching 62%.
The change suggests that model-selection decisions are becoming increasingly dynamic. Developers may be willing to switch between models based on price, latency, capability and workload rather than remaining tied to one provider.
This could make the AI market more competitive because model providers must continuously demonstrate value rather than relying solely on brand recognition.
DeepSeek’s Pricing Strategy Adds To Its Appeal
DeepSeek has built a reputation around offering powerful models at relatively low inference costs.
Recent pricing data for its V4 family shows substantial differences between peak and off-peak API pricing, while its broader strategy has emphasized efficient models and open-weight availability. Third-party pricing references based on DeepSeek’s published rates show V4-Flash and V4-Pro available at substantially lower prices than many leading proprietary systems.
The economic advantage becomes particularly relevant for applications where models are called repeatedly.
For example, an AI coding agent may need to process thousands or millions of tokens during development. Similarly, customer-service agents, research systems and software-development tools can produce large recurring inference bills.
US Platforms Are Becoming A Battleground For Chinese AI
The Vercel data is significant because the adoption is taking place on a U.S.-based developer platform.
This means Chinese AI models are not simply gaining users within China or on Chinese technology platforms. They are increasingly being incorporated into development workflows that use U.S. infrastructure.
That creates a new dimension in the technology competition between the two countries.
The United States continues to have major advantages in AI chips, cloud infrastructure and frontier proprietary models. At the same time, Chinese companies are demonstrating that model efficiency, lower costs and open-weight distribution can provide another route to international adoption.
The Hardware Restrictions Equation
Chinese AI developers continue to face constraints around access to the most advanced AI accelerators because of U.S. export controls.
That makes efficiency particularly valuable.
If a model can deliver competitive performance using fewer computational resources, its developer may be able to compensate partially for limitations in access to leading-edge hardware.
The rise of DeepSeek and other Chinese models therefore has implications beyond software. It is also becoming part of a broader debate over whether algorithmic efficiency can offset some of the advantages created by access to the most advanced chips.
Open Models Could Pressure US AI Companies On Pricing
The growing usage of Chinese open-weight models could put pressure on U.S. AI providers to lower prices or improve performance.
The competitive equation is no longer simply:
Which model is smartest?
It is increasingly:
Which model delivers the best combination of intelligence, speed, reliability and cost for a specific workload?
For production applications, particularly agents, cost can be decisive.
A developer may prefer a frontier proprietary model for difficult reasoning while choosing DeepSeek, StepFun or Z.ai for routine tasks. This could lead to increasingly sophisticated multi-model architectures rather than a single-model market.
The Bigger Picture
DeepSeek’s rise on Vercel’s AI Gateway highlights a major change in the global AI market: developers are increasingly treating model cost and deployment flexibility as core competitive factors. Open-weight models accounted for 54% of token volume on the platform on August 25 and reached 62% three days earlier, reversing their 28% share in June.
The broader significance is that Chinese AI models are gaining traction on a U.S. developer platform despite the wider technology rivalry between the two countries. DeepSeek is leading the shift, but StepFun and Z.ai are also gaining positions among the most-used models. If this trend persists, the AI market could become increasingly fragmented, with developers choosing models dynamically according to price, capability and workload.
Looking Ahead
The next test for Chinese open-weight models will be whether their recent surge can translate into sustained production adoption. Developers may continue using lower-cost models for high-volume agentic workloads while reserving premium proprietary systems for applications that demand the highest reasoning performance. That would create a more diversified AI ecosystem in which multiple model providers compete on different dimensions rather than one company dominating every workload.
For U.S. AI companies, the trend could increase pressure to improve token economics and make their models more flexible for developers. For Chinese AI companies, continued overseas adoption would demonstrate that open-weight distribution and cost efficiency can overcome some of the barriers created by geopolitical and hardware restrictions. DeepSeek’s position at the center of the shift makes its future model releases and pricing decisions particularly important for the next phase of the AI competition.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
