AI Market: The Metric of Success is the Dollar, Not the Token
Managing partner of venture firm Dragonfly, Haseeb Qureshi, has sharply criticized current practices for evaluating the AI model market. His main thesis: token consumption share is a misleading and often erroneous metric. In his opinion, the only correct way to compare is to analyze monetary costs, not raw data on the number of tokens.
Four Problems with the Token Metric
Qureshi highlights four systemic distortions that make analysis based on tokens on the OpenRouter platform practically useless for assessing real market position.
1. Subsidies and "free lunch". Chinese labs regularly launch new models with aggressive discounts or free access. This attracts users who migrate from one free model to another, creating an illusion of rapid token consumption growth but generating no real monetary revenue.
2. Model size. Small models, such as Qwen 3.5-27B, are about a hundred times cheaper per token than flagship Claude Opus or GPT-5.5 Pro. An increase in their usage might look like a sharp jump in market share on a chart, even though it is economically insignificant. According to Qureshi, the market should be analyzed strictly within weight categories by model size.
3. Multi-agent systems. For the same amount of money, one can run a complex multi-agent architecture based on DeepSeek, which will "burn" many times more tokens than a single request to the premium Opus model. Qureshi provides a clear example: if 5% of Opus usage shifts to such a system with a fourfold token consumption, the chart would show an 18% loss of Opus share, while the client's actual spending would only shift by 5%. This is a gross distortion of the picture.
4. Platform sampling. OpenRouter is not a universally representative cross-section of the market. Large companies, once they decide on a lab, prefer to work directly with Anthropic or OpenAI, bypassing the platform's markup. On the chart, this looks like a decline in the share of American models, even though the tokens simply move off the platform. OpenRouter is useful for analyzing share within the open model segment but is completely unsuitable for comparing open and closed models.
Price Pressure and the Future of Cheap Models
This point of view is developed by SageRoad Research founder Trevor Noren, linking it to intense price pressure on the industry. He cites a JPMorgan estimate that Amazon already offers about half of open models at a price that is only a fraction of the cost of flagship models. Nvidia, together with Dell, Lenovo, and HP, is creating computers for AI agents, which also contributes to shifting demand.
The most striking example is the price-to-performance ratio. Running the Artificial Analysis Intelligence Index task set on Claude Opus 4.8 costs $3,700 for a result of 56 points. DeepSeek V4 Pro scores 44 points for just $186 — roughly 20 times cheaper. Noren's conclusion is unequivocal: top-tier intelligence is not needed for everything. For the vast majority of corporate tasks, cheap, narrowly specialized models are sufficient, and cost control will be the main driver of this shift.
Expert opinion: The arguments of Qureshi and Noren hit the nail on the head. The AI market is overheated with vanity metrics. Investors and analysts who continue to look at "raw" tokens risk missing a tectonic shift towards efficiency and savings. The real battle in AI will not be over the best benchmark, but over the best price-to-performance ratio. Chinese labs DeepSeek and Alibaba are already winning this war, and Western giants will either have to radically cut prices or lose ground in the corporate segment.