DeepSeek’s recent decision to significantly reduce prices on its V4-Pro model by 75% should have been unequivocally good news for marketers and enterprise AI developers. Instead, many find that cheaper models don’t automatically translate into healthier margins.
The reason is simple: While consideration costs are falling, agent systems are consuming tokens faster than prices are falling. Two decades ago, the economics of software were dictated by the same rule. Infra became cheaper every year and the demands became stronger. AI was initially thought to follow a similar pattern. As frontier models improved and token prices fell, many assumed that consideration would be a negligible operational cost. That assumption has quickly begun to crumble.
A chatbot usually turns a single user’s query into a single model call. The agent turns it into a series of planning, retrieval, tooling, validation, summarization, and follow-up decisions. The user sees one answer. The seller pays for the loop. That’s a 100x problem: The same user interface request can be more expensive to execute as an agent workflow than a chatbot response or retrieval-augmented generation (RAG). For long-running jobs, the multiplier is high. Falling model prices help, but they don’t fix a product structure that turns one input into multiple billable functions.
A measure of what is at stake is now evident in the way the model providers themselves value developer relationships. OpenAI’s proposed plan to give all Y Combinator startups $2 million in API credits – a number that would have funded a seed round in any previous technology round, and where the same group benefits from a few thousand dollars in AWS credits – is less of a rent gain than the adoption of what it costs now to run an AI company in the first year of production. For agents returning established businesses to existing product lines, the absolute numbers are even larger.
What is token raising
In a single conversation chatbot, a single user message generates approximately one model call. The input-to-charge ratio is about 1:5.
For a multi-step agent outsourced from all customer support, sales, finance, legal review, and engineering tasks, that average often comes to 1:700 or more. Each loop iteration forwards the accumulated dialog, tool output, and logic trace. Each step adds up; it doesn’t lose anything.
A "simple" agent query like “What did our top customers ask about last week?” usually involves seven price functions before returning the answer:
-
User information (~50 tokens)
-
System definitions and tool definitions (~3,000 tokens, repeated for each call)
-
Recovery (~5,000 context tokens)
-
Model call #1 – tool selection (8,000 in / 200 out)
-
Using the tool (~4,000 tokens returned)
-
Telephone model #2 — summary (12,000 in / 400 out)
-
Model call #3 — tracking resolution (12,400 in / 100 out)
One sentence in, about 35,000 input tokens charged. Somewhere between $0.10 and $0.40 per query in the frontier model. Multiply that by a million inquiries per month – table capacity for any B2B aspect of the business – and the line item is six figures.
Why this breaks the existing AI business model
The story of the leading price of business AI has been chair-based SaaS: Pay per user per month, bring agent power, capture limit. That model assumes a reasonably limited cost per user.
Raising tokens breaks the assumption. A power user running 50 agent requests per day on a $40/seat plan can cost more in terms of understanding than the cost of the plan. Tokenization breaks the traditional SaaS pricing model. If the daily activity of a powerful agent user costs more than their monthly subscription fee, the seller’s gross margins turn negative, a compounding paradox as customers deepen the agent’s acquisition, the same consumption curve sellers sell their boards. Several vendors are now privately reporting negative gross margins for heavy users, simulating recent cloud cost reports from the Bessemer ‘Supernova’ group, where the correlation between AI agent adoption and gross margin contraction has gone from a belief risk to a major P&L storm.
Visible signs have begun to emerge. Bloomberg this week documented the growing gap between Salesforce’s Agentforce sales demos and customer deployment capabilities. This is the type of gap that opens up predictably where the promised performance is technically possible but it is not financially feasible to price the specified seat plan. Salesforce is the most watched case, not the exception.
"In my team, accounting costs are more than labor costs." – Bryan Catanzaro, VP of Applied Deep Learning, Nvidia
The purpose of the strategy does not exist "AI is expensive." That the dominant business model envisioned by many native AI company programs does not survive in relation to agency workloads.
A simple example
Consider a business software vendor that charges $40 per user per month for an AI-enabled support assistant. A typical chatbot might cost only a few cents per user per day per day for insight, leaving healthy gross margins.
Now replace that chatbot with a complete workflow capable of investigating tickets, querying internal systems, writing responses, validating outputs, and increasing diversity. If a heavy user makes 50 to 100 agent requests per day, the usage of the definition can increase by an order of magnitude. What were once neglected infrastructure costs are becoming operational costs.
This creates an unusual dynamic: The customers who get the most value from the product are often the customers who generate the highest perceived costs. In the worst cases, marketers can find themselves with highly engaged users offering little profit. The result is a growing realization across enterprise software that agent acquisition and margin expansion are no longer automatically aligned.
Agent orchestration is a new channel
Technical responses are known and convergent. They are not novels, but they are essential for survival
-
A cost-conscious route: This process includes a small classification model that determines which level (Haiku, Sonnet, Opus equivalent) handles each question. Well-configured routers cut projection bills by as much as 60 percent without quality degradation
-
Fast caching: Anthropic, OpenAI, and Google are now offering 75 to 90% discounts on archived startups.
-
Context discipline: You can reduce tool output, prune logic traces, and cap tool depth to prevent your agent from going down the rabbit hole.
-
Predictive coding: in a self-hosted implementation, this process ensures 2 to 3X efficiency on the same GPUs.
"Organizations using orchestration-led governance report strong productivity gains – a comprehensive orchestration layer is associated with a six-fold greater productivity impact than compliance-only approaches" – IBM
Companies that build this layer well are starting to look less like microservice operators and more financial trading systems: Every route decision has a price, every route has its P&L, every tenant has a limited budget.
What business leaders really need to do
Four move separates companies that will still have a margin in 24 months from those that won’t:
-
Make assumptions a first-class metric. It tracked per feature, per tenant, per question category the same way cloud spending has been tracked since the mid-2010s.
-
Budget as a media buyer. Set a ceiling of 1,000 cost inquiries per feature. They covered him. A warning on the run. Engineering cannot enforce this alone.
-
Treat the router as a core infrastructure, not a configuration. New load balancer.
-
Audits are reported quarterly. A 4,000 token program that has grown organically over the course of six months is a slow six-figure building. Most teams never read their own productivity alerts in the end.
-
The volume of the discussion is done ahead of time. Frontier model dealers are now offering premium style pre-paid jobs at huge discounts. The list price is the worst price any company will ever pay.
In the next 24 months
Structural change under agent AI is not that expensive. As DeepSeek’s price cut today emphasizes, the cost per unit of measurement on the border is falling by about 3X per year, and the curve is not flattening.
Change is what it is amplification outweighs price reduction. Cutting each token cost by 75% doesn’t help a company whose agents make 700X more tokens per user query than their assumed pricing model. For the first time since the dawn of the cloud era, architectural decisions are also financial decisions in real time. Rapid redesign is a margin event. The loop for an agent that isn’t tied properly is to turn it off with an attached credit card.
The companies that will survive the next 24 months of AI infrastructure pricing won’t be the ones that use the cheapest model. It will be those who work for wisdom again they know what it costs to think.
That’s a 100X problem. And it’s coming faster than the price cuts can hide.
Maitreyi Chatterjee is a senior software engineer at a large technology company.
Devansh Agarwal works as an ML engineer in a leading technology company.
