Dollars, Not Tokens: Why the Market Share of AI Models Is a Dangerous Illusion
Dragonfly Managing Partner Haseeb Qureshi has sharply criticized the current approach to evaluating the AI model market. In his view, analysis based on token consumption—especially on the OpenRouter platform—leads to fundamentally flawed conclusions. The real metric is dollars spent by users, not the volume of raw tokens.
The problem, according to Qureshi, lies in several key distortions. The first is subsidies. Chinese labs regularly launch models with aggressive discounts or even free access. This attracts a stream of users who jump from one free model to another, inflating token statistics but generating no real revenue. The second factor is the huge disparity in token cost depending on model size. A small model like Qwen 3.5-27B can be a hundred times cheaper per token than the flagship Claude Opus. A sharp spike in Qwen's share on the OpenRouter chart looks impressive, but economically it may be a drop in the bucket. Correct analysis, Qureshi emphasizes, is only possible within weight classes.
Token-Consuming Systems and OpenRouter's "Blind" Statistics
The third trap is multi-agent systems. You can spend the same amount on a complex architecture based on DeepSeek or GLM 5.2 as on a single query to a top-tier model like Opus. With comparable performance, a multi-agent configuration will burn many times more tokens for the same money. "If 5% of Opus usage shifts to such a system with a fourfold token expenditure, the chart will show an 18% loss of Opus's share, even though actual spending shifts by only 5%," Qureshi explains. This dramatically exaggerates the importance of cheap tokens.
The fourth nuance is the very choice of OpenRouter as a data source. If a company has already settled on a provider (Anthropic or OpenAI), it is more profitable to work directly, bypassing the platform's markup. On OpenRouter charts, this looks like a decline in the share of American models, even though real consumption simply moves off the platform. Qureshi's conclusion is uncompromising: OpenRouter is useful only for comparisons within the open model segment, but is completely unsuitable for comparing open and closed ecosystems.
Price Pressure: Is the Future in Cheap Models?
This logic is developed by SageRoad Research founder Trevor Noren. He cites JPMorgan data: many future tokens will be consumed not by cutting-edge models, but by small open models sufficient for specific tasks. Amazon already offers about half a thousand open models at a fraction of the cost of flagships, while Nvidia, together with Dell, Lenovo, and HP, is creating computers for AI agents. Meanwhile, the American giants' own mini-models (Claude Haiku, GPT-5.4-mini) are currently uncompetitive on the "efficient frontier," where Chinese developers—DeepSeek, MiniMax, Xiaomi, and Alibaba—now dominate.
The cost example is particularly telling. Running the Artificial Analysis Intelligence Index benchmark on Claude Opus 4.8 costs $3,700 for a score of 56. DeepSeek V4 Pro scores 44 points for just $186—roughly 20 times cheaper. The conclusion is obvious: cutting-edge intelligence is not needed for everything, and where it is necessary, GLM 5.2 from Z.ai is already comparable to Anthropic and OpenAI's top models.
Expert opinion: The arguments of Qureshi and Noren are not just an academic debate. For the crypto industry, where "on-chain" metrics are often perceived as absolute truth, this is an important reminder: raw data without the context of monetization can be misleading. The AI market is moving toward commoditization, and the real battle will be not over tokens, but over the dollars of corporate budgets seeking maximum efficiency. The victory of Chinese models in the "price-quality" segment is not just a trend, but a long-term structural shift.