Open-source AI models gain ground in 2026
Open source AI models are increasingly gaining ground in the business environment, amidst cost reductions, rising competition, and the need for flexibility. According to the Vercel AI Gateway Production Index – October 2026 report, models with open parameters accounted for 71.5% of the token volume processed through the Vercel AI Gateway in September 2026. During the same period, they generated 23% of the expenses recorded through the platform.
Data published by Vercel indicates a significant shift in the artificial intelligence market: companies are using more models, comparing providers, and distributing their budgets among solutions with different prices and features.
The report also highlights another relevant trend: while total spending on artificial intelligence models has increased sharply, the average price for a token has dropped by 47% compared to December 2025.
Open source and open-weights AI models dominate usage volume
One of the most important findings of the Vercel report is the 71.5% share of the total token volume processed through the AI Gateway attributed to models with open parameter access in September 2026.
A token is a unit of text used by artificial intelligence models for processing information. The volume of tokens can include text input by the user, generated responses, tokens used for reasoning, and certain data reused through temporary memory mechanisms.
According to Vercel, models with open parameter access surpassed restricted access models for the first time in terms of processed volume in August 2026. In September, this trend solidified.
However, a technical distinction is important: the term “open source” is not always equivalent to “open parameter access.” Some models allow access to the weights learned during training, without making the training data and all components needed to reproduce the process publicly available.
Therefore, the Vercel report analyzes models categorized as having open parameter access, according to the classification used by the platform. This category should not be automatically confused with the strict definition of a completely free model.
Spending on AI models with open parameter access has increased over 70 times
The increase in usage is accompanied by a significant change in budgets allocated to artificial intelligence. In June 2026, open-weights models represented about 5% of spending through Vercel AI Gateway. In September, their share had reached 23%, meaning that nearly one in four dollars spent through the platform was allocated to these models.
Furthermore, Vercel shows that, from December 2025 to September 2026, spending on models with open parameter access increased over 70 times. During the same period, spending on top models with restricted access grew by about ten times.
These figures do not necessarily indicate that open source AI models completely replace commercial solutions. On the contrary, the report shows that many teams are simultaneously increasing their budgets for both categories.
The trend suggests that organizations choose models based on the task they need to perform, cost, and results obtained, rather than relying exclusively on a single provider.
DeepSeek demonstrates that high processing volume does not automatically imply high costs
A relevant example from the report is DeepSeek-V4.1-Flash, which processed nearly half of the total tokens recorded in September through Vercel AI Gateway, while representing only 2.6% of the expenses associated with the category of models that offer open parameter access.
According to Vercel, the price per token for this model was approximately five times lower than the category average.
This result highlights an important feature of the market: low-cost models can process very large volumes of information without generating a similar share of expenses.
For companies, such solutions can be appealing in activities such as document classification, data extraction, information summarization, and automating repetitive tasks.
However, cost should not be the only selection criterion. The quality of responses, processing time, security, and compatibility with existing applications must be evaluated before implementation.
Kimi and DeepSeek illustrate two different strategies
The Vercel report shows that open-weights models contribute to increased spending across various price segments.
In September 2026, Kimi K3 models, developed by Moonshot AI, made a significant contribution to the increase in spending for this category. Vercel estimates that models with a cost per token at least four times higher than the category average generated approximately two-thirds of the increase in spending for models with open parameters.
In contrast, DeepSeek-V4.1-Flash attracted a very high processing volume at a relatively low cost.
This difference shows that open parameter models do not form a uniform category. Some are used for high-volume economic requests, while others may attract larger budgets due to their features and usage methods.
For companies, the right choice depends on the balance between performance, cost, and the complexity of the activities for which the model is used.
Costs are decreasing, AI usage is increasing
Another important result of the report is the approximately 13-fold increase in total spending through Vercel AI Gateway in September 2026, compared to December 2025.
During the same period, the estimated average price per token dropped by 47%. In other words, organizations are spending more on artificial intelligence, even though the average processing price has become lower. One explanation is the expansion of usage: companies are integrating these technologies into more applications and processes.
Vercel also shows that a median paying client, active in both December and September, more than doubled their spending during this period. This trend suggests that lowering the unit cost can encourage increased usage. When processing becomes more affordable, organizations can automate additional activities or use models more frequently.
However, the decrease in average price does not guarantee a reduction in total budget. Spending also depends on the volume of requests, the complexity of tasks, and the models chosen.
Companies are increasingly changing AI models and providers
The Vercel report also highlights increased mobility among teams using artificial intelligence. In September 2026, nearly one in two teams changed the most frequently used model, and nearly one in three changed their preferred model provider.
For the second indicator, Vercel reported a record within its index. Organizations that processed at least one billion tokens from December 2025 to September 2026 used, on average, 42 different models, counting each version separately.
This data shows that many companies no longer rely on a single solution for all activities. They can use different models for different tasks and can change configurations when more efficient alternatives arise. For organizations developing AI-based applications, this flexibility can reduce dependence on a single provider. At the same time, it requires testing, monitoring, and verifying compatibility between models.
Chinese AI models are also used through American providers
The Vercel report also presents an interesting trend regarding the commercial distribution of models developed in China. In September 2026, nearly 70% of spending through the platform for Chinese models was directed through providers based in the United States. In December 2025, this share was less than 10%.
For example, DeepSeek models accounted for 58% of the total token volume processed through the platform, but 6.6% of expenses. Over four-fifths of the tokens associated with DeepSeek models were processed directly by the company, while two out of five dollars spent on these models went through American providers. In the case of Moonshot AI, the company behind Kimi models, American providers processed approximately nine out of ten tokens associated with its models through the analyzed platform.
This data shows that developing a model and distributing access to it are distinct activities. A model can be created by a company in one country and offered to users through the commercial infrastructure of a company in another country.
For organizations, this aspect highlights the importance of verifying the actual provider of the service, the conditions of data processing, and the applicable legal requirements.
What should companies analyze before adopting open source AI models?
The growing popularity of open source models and those with open parameters does not mean they are suitable for every application. Before implementation, companies should evaluate:
- Total cost: usage price, infrastructure, administration, and maintenance.
- Quality of results: accuracy of responses relative to actual tasks.
- License: rights for commercial use, modification, and distribution.
- Security: data protection, access control, and management of confidential information.
- Flexibility: the ability to change the model without major changes to the application.
- Technical resources: equipment and skills necessary for managing the solution.
In particular, organizations should avoid choosing a model solely based on popularity or price. A cheap solution can become costly if it requires frequent correction of results or additional infrastructure.
Therefore, the strategic advantage lies not only in choosing an open source model but in the ability to select, compare, and efficiently use multiple models.
