The true cost of AI models: why the dollar matters more than tokens
Market analysis of AI models based on token consumption is a path to fundamentally flawed conclusions. Leading industry experts have reached this conclusion, pointing out that the real measure of competitiveness and demand should be user dollar spending, not raw token usage metrics.
Problem #1: Data distortion by subsidies
Chinese labs are actively using aggressive pricing strategies, launching new models with huge discounts or even free access. This creates an illusion of high demand: users switch between free models, inflating token consumption statistics, but without spending real money. Comparing such activity with paid usage of flagship models is inherently misleading.
Problem #2: Model size matters
Small models, such as Qwen 3.5-27B, can be a hundred times cheaper per token than, say, Claude Opus. A sharp spike in the usage share of cheap models on OpenRouter charts looks like a triumph of open AI, but in monetary terms it may be completely insignificant. The market must be analyzed strictly within weight categories, comparing models of comparable size.
Problem #3: Multi-agent systems
Modern architectures using cascades of multiple models (e.g., based on DeepSeek) burn far more tokens for the same money than a single powerful model. If 5% of Opus usage shifts to such a system with four times the token consumption, the chart will show an 18% loss in Opus share, while actual dollar spending drops only 5%. Tokens exaggerate the significance of cheap solutions.
Problem #4: Limitations of the OpenRouter platform
Large corporations, once they have chosen a provider, prefer to work directly with Anthropic or OpenAI, bypassing OpenRouter's markup. This means that a decline in the share of American models on the platform does not reflect real customer churn — tokens simply move beyond OpenRouter's visibility. The platform is useful for assessing trends within the open model segment but is completely unsuitable for comparing open and closed ecosystems.
Is the future in cheap models?
JPMorgan provides a striking example: running a benchmark on Claude Opus 4.8 costs $3,700 (result: 56 points), while DeepSeek V4 Pro achieves 44 points for just $186 — 20 times cheaper. Amazon already offers half a dozen open models at a fraction of the cost of proprietary counterparts, and Nvidia, together with Dell, Lenovo, and HP, is creating computers for AI agents. Corporate spending remains the main driver of AI investment returns, but companies will minimize costs.
My conclusion: The industry is moving toward commoditization. Leading positions on the "efficient frontier" are currently held by Chinese developers — DeepSeek, MiniMax, Xiaomi, and Alibaba. Measuring the market in dollars, not tokens, will inevitably shift the focus of analysis toward cheap, highly specialized models. And this is not a question of the future — it is happening right now.