AI agents’ token usage surge makes seat‑based SaaS pricing untenable
DeepSeek's 75% price cut on its V4-Pro model makes inference cheaper at the per-token level, but the source frames this as insufficient against a structural problem: agent workflows amplify token usage from roughly 1:5 in a single-turn chatbot to 1:700 or higher in multi-step agents. The same user-visible request can cost vastly more to serve through an agent than through a chat response, and the seat-based SaaS model most enterprise AI vendors still rely on does not survive that shift.
The amplification mechanism is concrete. A chatbot turns one user question into one model call. An agent turns the same question into a chain of planning, retrieval, tool use, verification, summarization, and follow-up decisions. The source walks through a "simple" agent query ("What did our top customer ask about last week?") that the article says touches seven priced operations: a roughly 50-token user prompt, a 3,000-token system prompt and tool-definition block repeated on every call, 5,000 tokens of retrieval context, a first model call for tool selection at 8,000 input and 200 output tokens, a 4,000-token tool execution result, a second model call for summarization at 12,000 input and 400 output tokens, and a third for follow-up decision at 12,400 input and 100 output tokens. One sentence in, the source says, around 35,000 input tokens are billed.
At frontier-model pricing, the source puts that single query between $0.10 and $0.40. Multiplied by what the source calls the table-stakes enterprise volume (a million queries a month for any B2B feature), the line item lands in six figures. The implication is not that AI is expensive; it is that the dominant business model assumed by most AI-native plans does not survive contact with agentic workloads.
The economic pressure shows up in the gap between subscription revenue and per-user inference consumption. A heavy user running 50 to 100 agent invocations per day on a $40 per-seat plan can cost more in inference than the plan charges. The source cites unnamed vendors as privately reporting negative gross margins on heavy users and points to a pattern in the Bessemer "Supernova" cohort where AI-agent adoption correlates with gross-margin contraction. The customers receiving the most value from the product become the customers generating the highest inference costs, in what the source frames as a paradox that compounds as customers deepen agent adoption.
The visible symptoms have started leaking into public coverage. The source cites this week's Bloomberg reporting on a widening gap between Salesforce's Agentforce marketing demos and what is actually shipping to customers. The source reads this gap as a predictable outcome when promised functionality is technically possible but uneconomical to serve at the seat price. Salesforce is the most-watched case, the source argues, not a unique one.
The pressure shows up at the model-provider level too. The source quotes Nvidia's Bryan Catanzaro, the company's VP of Applied Deep Learning, as saying "the cost of compute is far beyond the costs of the employees" for his team. The same source reports that OpenAI has proposed giving every Y Combinator startup $2 million in API credits, which the source describes as less a recruiting perk than an admission of what it now costs to run an AI-native company through its first year of product.
The technical responses are converging rather than novel. Cost-aware routing, where a small classifier picks the model tier for each query, the source says, cuts inference bills by around 60% without quality loss. Prompt caching, the source reports, now offers 75 to 90% discounts on cached prefixes from Anthropic, OpenAI, and Google. Context discipline, in the source's framing, includes truncating tool outputs, pruning reasoning traces, and capping tool depth to prevent runaway loops. In self-hosted deployments, speculative decoding yields 2 to 3x effective throughput on the same GPUs. None of these are new; what is new is treating the orchestration layer as core infrastructure rather than optimization.
The orchestration framing extends beyond engineering. The source cites IBM as saying that organizations using orchestration-led governance report stronger productivity gains, and associates a holistic orchestration layer with "six times greater productivity impact than compliance-only approaches." Companies that build this layer well, the source argues, are starting to look less like microservice operators and more like financial trading systems: every routing decision priced, every path with its own P&L, every tenant on a metered budget.
The source's prescription for the next 24 months: treat inference cost as a first-class metric tracked per feature, per tenant, and per query class, much as cloud cost was tracked starting in the mid-2010s; budget like a media buyer with cost-per-thousand-query ceilings and overruns that alert; treat the router as core infrastructure rather than an optimization layer; audit prompts quarterly, since a system prompt that grew organically over six months can quietly become a six-figure bill; and negotiate volume commits early rather than paying list, which the source frames as the worst price any enterprise will ever pay.
The companies that survive the next 24 months of agent pricing pressure will be the ones whose agents know what they cost to think, not the ones running the cheapest model. The orchestration savings the source names (around 60% from routing, 75 to 90% from cached-prefix discounts, 2 to 3x from speculative decoding) are presented as known industry practice, but the source does not report how those savings compose in a single production stack or whether they hold under concurrent load. The Salesforce Agentforce coverage the source cites documents one visible instance of the price-versus-capability gap, not a systematic one, and whether the 100x problem is the new P&L headline for every AI-native vendor or only for those running the most ambitious agents is the gap the source does not close.