DeepSeek’s latest decision drastically reduced prices its 75% in the V4-Pro model should definitely be good news for enterprise AI vendors and developers. Conversely, many are discovering that cheaper models don’t automatically translate into healthier margins.
The reason is simple: Agent systems are greedily consuming tokens faster than prices are falling, even though the resulting costs have dropped dramatically. For the past 2 decades, the software economy has been dictated by the same rule. Infrastructure has become cheaper every year, and applications have become more capable. Artificial intelligence was initially assumed to follow the same pattern. As frontier models improve and token prices fall, a lot of speculation will translate into negligible transaction costs. That assumption began to disintegrate exponentially.
A chatbot usually turns a user question into a model call. Agent making it a chain of planning, searching, using tools, checking, summarizing, and subsequent decisions. The user sees a response. The seller pays for the loop. This is a 100x problem: The same query seen by a user can cost more than a chatbot or search augmented generation (RAG) response to serve as an agent workflow. In longer-term workflows, the multiplier is higher. Drop-down model pricing helps, but doesn’t fix product architecture that turns one order into dozens of billable transactions.
The extent of what is now at stake is clear in how the model providers themselves value their developer relationships. OpenAI’s proposed program of giving each Y Combinator startup $2 million in API credits — a number that would have funded an entire seed round in any previous tech round, and when the same cohort received several thousand dollars in AWS credits — is less of a recruiting bonus than accepting what it costs to run an AI-powered company in its first year. The absolute numbers are still huge for established businesses upgrading agents to existing product lines.
What is token boosting?
In a single-turn chatbot, one user message generates approximately one model call. The included texture ratio is about 1:5.
a multistep agent spread across customer support, sales operations, finance, legal review, and engineering, the ratio is steadily declining 1:700 or higher. Each loop iteration advances the cumulative conversation, tool results, and reasoning traces. Each step adds up; nothing falls.
A "simple" such as agent requestWhat did our best customer ask last week? before returning a response, it typically performs seven value operations:
-
User request (~50 tokens)
-
System command and tool definitions (~3000 tokens, repeated per call)
-
Search (~5000 context clues)
-
Model challenge #1 — tool selection (8000 in / 200 out)
-
Tool execution (~4000 tokens returned)
-
Model challenge #2 – summary (12,000 in / 400 out)
-
Model call #3 — chase decision (12,400 in / 100 out)
In one sentence, approximately 35,000 access tokens are counted. It is between $0.10 and $0.40 per request in the frontier model. Multiply that by one million queries per month—the amount of table shares for any enterprise B2B function—and the line item is six figures.
Why does this disrupt the existing AI business model?
Dominant price story for enterprise AI has been seat-based SaaS: Pay per user monthly, deliver agent capability, capture margin. This model assumes a reasonably limited cost for each user.
Token amplification violates the assumption. A power user with 50 agent calls per day on the $40/seat plan can get more results than the plan charges. Token empowerment disrupts the traditional SaaS pricing model. When a power user’s daily agent activity costs more than the monthly subscription fee, the seller’s gross margin is negative, a paradox that results in customers deepening agent adoption, the usage curve selling sellers to their boards. Several vendors are now privately reporting negative gross margins for heavy users, mirroring recent cloud spending reports from the Bessemer ‘Supernova’ cohort, where the correlation between AI-agent adoption and gross margin contraction has shifted from theoretical risk to major P&L headwinds.
Visible symptoms began to leak into the public domain. Bloomberg documented this week the widening gap between Salesforce’s Agentforce marketing demos and its ability to actually deliver to customers. This is the type of gap that can be predicted when the promised functionality is technically possible but uneconomical to serve at the cost of the seating plan. Salesforce is the most watched event, not rare.
"For my team, the cost of computing far outweighs the cost of employees." — Bryan Catanzaro, VP of Applied Deep Learning, Nvidia
It has no strategic meaning "AI is expensive." The dominant business model adopted by most AI-native company plans does not survive contact with agent workloads.
A simple example
Consider an enterprise software vendor that charges $40 per user per month for an AI-powered support assistant. A traditional chatbot can ultimately cost only a few cents per day per user, leaving healthy gross margins.
Now replace that chatbot with a full agent workflow capable of investigating tickets, querying internal systems, compiling responses, validating results, and raising exceptions. If a heavy user performs 50-100 agent requests per day, the resulting consumption can increase by orders of magnitude. What was once a modest infrastructure cost becomes a significant operational cost.
This creates an unusual dynamic: The customers who get the most value from the product are often the customers with the highest results. In extreme cases, sellers may find themselves with the most engaged users making the least profit. The result is the growing realization across enterprise software that agent adoption and margin expansion are no longer automatically matched.
Agent orchestration is the new ditch
The technical answers are known and convergent. They are not new, but they are critical to survival
-
Cost-aware routing: This technique involves a small classification model that decides which level (Haiku, Sonnet, Opus equivalents) should handle each query. Well-tuned routers reduce inference bills by about 60% without any loss in quality
-
Fast caching: anthropicOpenAI and Google are now offering 75-90% off cached prefixes.
-
Context discipline: To prevent your agent from going down the rabbit hole, you can cut tool exits, ground traces, and cap tool depth.
-
Speculative decoding: for self-hosted deployments, this technique guarantees 2-3X effective throughput on the same GPUs.
"Organizations using orchestration-based management report stronger productivity gains—a single orchestration layer has six times greater productivity impact than compliance-only approaches." — IBM
Companies that build this layer well start to look less like microservices operators financial trading systems: Every route decisions priced, every route with its own P&L, every tenant on a measured budget.
What should business leaders actually do?
FOur actions separate companies that will still have margins in 24 months from those that will not:
-
Make the result value a first-order metric. Look at it per feature, per tenant, per request class, as cloud spending has been tracked since the mid-2010s.
-
Budget as a media buyer. Set cost per thousand requests ceilings for each feature. Cover them. Warning about exceeding the limit. Engineering will not implement it alone.
-
Treat the router as a core infrastructure, not an optimization. This is a new load balancer.
-
Quarterly inspection requirements. A 4,000-token system offering that has grown organically over six months is a six-figure bill in slow motion. Most teams never read their production instructions all the way through.
-
The volume of discussion is done early. Frontier model suppliers now offer protected sample style prepaid commitments at significant discounts. List price is the worst price any business will pay.
The next 24 months
Structural change under agent AI is not that it is expensive. As DeepSeek’s price cut today highlights, demarcation unit costs are falling roughly 3x per year, and the curve isn’t slowing.
That’s what change is all about reinforcement outweighs price reductions. A 75% reduction in cost per token doesn’t help a company whose agents make 700X more tokens for each user request than the assumed pricing model. For the first time since the dawn of the cloud era, architectural decisions are again real-time financial decisions. A quick redesign is a margin event. A poorly closed agent loop is a break where a credit card is closed.
The companies that survive the cost of AI infrastructure over the next 24 months will not be the ones running the cheapest model. Agents will be smart ones and they know what it costs to think.
This is a 100X problem. And it’s coming faster than price cuts can hide.
Maitreyi Chatterjee is a senior software engineer at a large technology company.
Devansh Agarwal works as an ML engineer at a leading tech company.
