DeepSeek announced a 75% reduction in the price of its V4‑Pro model, a move that should have eased inference costs for developers. In practice, many firms are finding that cheaper tokens do not translate into healthier profit margins because agentic AI systems consume tokens at rates far exceeding the price decline.
The article calls this the “100x problem”: a single user request routed through an agent can trigger dozens of model calls, each adding thousands of tokens. A typical multi‑step query—retrieving data, selecting tools, summarizing, and making follow‑up decisions—can bill roughly 35,000 input tokens, costing between $0.10 and $0.40 per query. At a scale of a million queries a month, the expense reaches six figures, far outpacing traditional seat‑based SaaS pricing.
Vendors are now treating inference cost as a core metric, deploying cost‑aware routing, prompt caching, and context discipline to curb token amplification. Enterprises are urged to monitor per‑feature usage, set query‑cost caps, audit prompts regularly, and negotiate volume‑commit discounts. The shift suggests that future AI profitability will hinge less on model price and more on disciplined agent architecture.
