What happened
The South China Morning Post reports that open-weight AI models accounted for 54% of token volume on Vercel’s AI Gateway on Tuesday, compared with 46% for proprietary models. The outlet says open-weight models reached a record 62% share on Saturday, citing Vercel data and comments from CEO Guillermo Rauch.
The South China Morning Post reports that open-weight models made up 54% of token volume on Vercel’s AI Gateway on Tuesday, overtaking proprietary models, which accounted for 46%. The article says the open-model share reached 62% on Saturday, compared with 38% for closed systems. SCMP attributes the Saturday figure to Vercel chief executive Guillermo Rauch’s social-media post and says the Tuesday figures came from data published on Vercel’s website. The report does not provide a separate audit of the platform’s measurements, and the figures have not been independently confirmed here.
The reported change reverses Vercel’s June 24 breakdown, when open-weight models represented 28% of token volume and closed models represented 72%, according to SCMP’s account of Rauch’s data. That comparison covers two points in time rather than a continuous series, so it establishes a sharp reported change but does not, by itself, explain when the shift occurred or whether it was steady.
SCMP reports that DeepSeek-V4-Flash was the most-used model by token volume as of Tuesday. Chinese models occupied four of the five leading positions: Step 3.7 Flash from StepFun ranked second, OpenAI’s GPT-5.6 Luna ranked third, GLM-5.2 from Zhipu, or Z.ai, ranked fourth, and an updated 0731 version of DeepSeek-V4-Flash ranked fifth. These rankings describe usage on Vercel’s platform, not overall global usage or model quality.
The article says the increased use of open-weight models was driven largely by DeepSeek’s latest lightweight model. SCMP also links the movement to business demand for Anthropic’s Fable 5 stalling because of high costs. That explanation is a report-specific claim, not an independently established finding in the supplied material; the source does not give adoption figures for Fable 5, customer interviews, pricing comparisons or evidence that cost was the only factor.
Read the primary source: scmp.com ↗
Why it matters
If the reported figures are representative, they show developers using cheaper open-weight models for production workloads that can consume large numbers of tokens. That could influence model pricing, infrastructure choices and competition between Chinese and US AI providers.
The practical significance is tied to the type of workload involved. SCMP reports that developers are choosing cheaper open models particularly for autonomous agents, which can repeatedly reason, write code and call software tools. Those repeated steps can generate substantially more token volume than a short one-off interaction, making per-token pricing and access to models that can be run or integrated flexibly important operational considerations.
The reported rankings also suggest that Chinese-developed models are becoming visible in a US-based developer infrastructure channel. That does not establish broader geopolitical or commercial dominance, but it does indicate that at least within the measured Vercel AI Gateway traffic, several Chinese open-weight systems were being used heavily enough to occupy most of the top five positions. For model providers, platform operators and enterprise buyers, usage in production-oriented tooling can matter more than attention generated by demonstrations or benchmark claims.
Open-weight access can give developers more control over deployment, model selection and cost, although the supplied report does not specify the licensing terms, hosting arrangements or technical constraints for any of the named models. “Open-weight” also does not necessarily mean fully open source: the article does not detail the models’ training data, software licenses, support obligations or ability to be modified. Those distinctions are important for organizations assessing legal, security and operational risks.
The figures could affect competitive pressure in the AI market if they persist across other platforms. A sustained movement toward lower-cost models would put pressure on providers whose products command higher prices, while increasing demand for routing, evaluation and monitoring tools that select among models. That is a possible industry implication, not a result demonstrated by the article. The source provides one platform’s usage data and does not include revenue, margin, customer-retention or quality measurements.
The report is therefore material as an indicator of model consumption and pricing pressure, but its scope should be kept narrow. It does not show that open-weight models outperform proprietary systems, that Chinese models are more capable, or that businesses generally have abandoned closed models. It shows that SCMP reports a major change in token-volume share on Vercel’s AI Gateway, with DeepSeek and other Chinese models prominent in that measurement.
What to watch next
The key unknowns are whether the shift persists, how Vercel measures token volume, and whether the data reflects a broad market trend or the behavior of Vercel’s developer customers. The South China Morning Post’s explanation linking the change to high costs for Anthropic’s Fable 5 also requires independent confirmation.
The first question is persistence. Future Vercel updates would show whether the 54% Tuesday share and 62% Saturday peak were temporary spikes or part of a continuing change. A longer time series would also clarify whether the June 24 level of 28% was an isolated baseline or representative of earlier traffic. Without that context, the magnitude and durability of the reversal remain uncertain.
The second question is measurement. Token volume is a useful indicator of model consumption, but it is not the same as the number of users, requests, applications, revenue or business value. The supplied report does not explain whether Vercel’s figures count input and output tokens in the same way, how traffic is attributed across model versions, or whether a small number of high-volume applications drove the result. Those methodological details would determine how broadly the figures can be interpreted.
The third question is model economics. SCMP attributes the shift partly to the high cost of Anthropic’s Fable 5, but the source does not provide prices, comparative usage costs or independent evidence of stalled business demand. Reporting from Vercel or customers about workload costs, model-routing decisions and performance requirements would help establish whether price, capability, availability, licensing or other factors drove the change.
The fourth question is deployment risk and suitability. The source says developers use these models for autonomous agents, code writing and software-tool calls, but it supplies no reliability, safety, privacy or security testing. Organizations considering similar systems would need evidence about error rates, data handling, license restrictions, uptime and human oversight before treating platform popularity as a reason to adopt a model.
Finally, broader corroboration matters. Usage data from other gateways, cloud providers, developer platforms or enterprise customers would indicate whether the reported Vercel pattern reflects a wider shift toward open-weight models. Until such evidence is available, the strongest supported conclusion is limited: SCMP reports a sharp and consequential change in model token share on one US platform, led by DeepSeek and other Chinese open-weight systems, but the trend’s cause and generality remain unresolved.
