Lower prices are changing how AI agents are built and run

• AI agents are driving token cost concerns behind OpenClaw’s Kimi K2.5 support.
• Lower prices ease spending pressure, with risks still under review.

In late 2025 and early 2026, a quiet shift has been underway in how AI services are being built and consumed. What once was a clear division—heavy reliance on expensive proprietary language models from Western firms—is now blurring. A growing number of users and developers are experimenting with open-source alternatives, especially ones emerging from China that cut costs sharply without a dramatic drop in performance.

This shift shows up in the recent support added by the autonomous AI agent OpenClaw for Chinese open-source models such as Moonshot AI’s new Kimi K2.5 and its coding variant. OpenClaw’s move suggests these models now sit low enough on the cost-performance curve to matter to a global audience, not just a niche of technical early adopters.

At the heart of this evolution are the economics of these models. AI service pricing is typically tied to token usage—the units of text a model processes and generates. Higher token volumes mean higher costs for users and developers. In the traditional proprietary model world, those prices can add up quickly for heavy use, especially when AI agents run autonomous tasks that generate large token counts. That’s one reason some users have reported unexpectedly large bills when letting agents operate without tight guardrails.

In contrast, Chinese open models built on transparent licences are attracting attention primarily for what analysts call “value for money.” For example, Moonshot’s Kimi K2.5 is offered at a fraction of the token cost charged by advanced Western models—around $0.58 for 1 million input tokens and $3 for output, which is roughly one-ninth and one-eighth of the pricing on some leading proprietary systems.

That gap isn’t an isolated case. Broader industry comparisons show open-source models available today can cost developers an order of magnitude less for equivalent token processing than many closed systems, without a proportionate drop in quality for many tasks.

This cost difference matters more than ever because enterprises and high-volume users are starting to weigh not just performance but sustainability. A recent study of open versus closed models found that, despite the lower price tags, closed models still represent about 80 % of global usage and 96 % of revenue, largely due to established brand trust and integration inertia. Analysts estimate that if adoption tilted toward open models based purely on performance and price, the global AI economy could save tens of billions of dollars annually.

OpenClaw’s choice to support Kimi and similar models reflects these pressures. While the founder of OpenClaw did not comment publicly on the decision, allowing users to access these models for free suggests a broader strategy: reduce cost barriers and broaden the agent’s appeal. Whether this approach improves long-term retention or monetisation remains an open question.

The technical capability of these models has improved in step with their price point. Moonshot’s Kimi K2 series, one of the more advanced Chinese open models, has been developing rapidly based on a mixture-of-experts architecture, high token context windows, and aggressive optimisation strategies.

Users of OpenClaw have varied in how they engage with the system. Some, like AI and machine learning professor Wang Shuyi in Tianjin, use it as a sort of “task router”: issuing research, report writingor creative directives at night and reviewing results the next day. Others lean on it for daily workflow augmentation, viewing the agent as a kind of virtual assistant.

What lower-cost models mean for AI agents and users

But the move toward open models isn’t just about price and performance. Companies and researchers caution that operational costs for open models aren’t limited to token pricing alone. Integrating, maintaining, and securing models within enterprise systems can require engineering investments that offset some of the cost advantage. One industry analysis notes that “open-source isn’t free” in a holistic sense and that businesses must account for infrastructure, maintenance, monitoring, and compliance costs.

That point resonates in user conversations beyond technical forums. Some OpenClaw users dismiss privacy concerns in favour of utilitywhile others hesitate or adopt workarounds like running services in the cloud due to fears over data security on personal systems. The balance of risk versus reward differs by user and use case.

China’s AI ecosystem has accelerated on multiple fronts, not just through Moonshot. Other domestic players, such as DeepSeek and ByteDance, have released low-cost models that have generated attention abroad for their performance per dollar of compute. These developments suggest that cost competition will remain a central theme in AI model adoption for the foreseeable future.

The broader takeaway is that the AI model market is fragmenting. For enterprise developers and high-volume users, the calculus is no longer simply “best model available.” It increasingly includes questions about price, flexibility, control, and long-term cost sustainability. As agents like OpenClaw embrace cheaper alternatives, that calculus will only become more visible—and more consequential for developers, businesses, and end users alike.

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology eventsclick here for more information.

AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.