OpenRouter's September 14 rankings show that seven Chinese models occupy the global top ten, representing about 72% of the tokens listed. This skew toward lower-cost models appears across other platforms as well: the largest agents' usage statistics are driven primarily by cost-based selection.
Specific ranking details
- On the Hermes agent's list, DeepSeek is ranked at the top.
- In OpenClaw, DeepSeek appears in both first and second place; together DeepSeek and GLM account for roughly 62% of OpenClaw's top-ten tokens.
- On Claude Code, the most-used model is GLM 5.3 Flash, used 4.4 times as often as Claude Opus 5.
These figures indicate that buyers of model tokens prioritize price per token when choosing which models to run.
Practical implications
Agents and services are buying on cost-efficiency rather than prestige: cheaper models that are “good enough” consistently win out in usage. This is not necessarily a market failure but a reflection of the market functioning under token-based incentives—cost competitiveness determines adoption.
As a result, frontier labs and leading AI teams, whose flagship models are more expensive to run, increasingly discuss slowing or pausing releases and cite safety concerns. Market data suggest the flagship models are losing ground not because of inferior capabilities but because they are less cost-effective under token-based purchasing.
Conclusion
In the current token-driven ecosystem, the central risk is not primarily that models are becoming demonstrably smarter, but that cheaper models proliferate. When the highest-performance products are also the least purchased due to price, discussions about safety and release pacing become the dominant public narrative.
