Inference is getting cheaper.
This is right in the same way that phone calls are cheap is right.
One phone call is cheap. A company with 10,000 employees calling, forwarding, checking, escalating and scheduling meetings all day is not.
That is where the AI economy is heading.
A chatbot is relatively easy to price. One person asks a question. One model produces an answer. Count the tokens, apply the rate and move on.
An agentic system is different. A planner breaks the task into pieces. A research agent opens documents. A second agent checks the research. A third challenges the assumptions. A supervisor asks for revisions. A memory layer retrieves old context. A compliance agent reviews the result. A final agent rewrites everything for the customer.
The user sees one answer.
Underneath it, a small synthetic company has held 60 meetings.
This is the hidden multiplier in the agent economy: the cost of intelligence may be falling, while the cost of coordination is exploding.
And because agent-to-agent communication happens behind the interface, most buyers will not notice until the invoice arrives or the gross margin disappears.
The cheap-token story is only half the story
The standard AI cost curve looks beautiful.
Models get smaller. Chips get faster. Quantisation improves. Caching gets better. Competition pushes token prices down. A task that cost a dollar eventually costs ten cents, then one cent.
This creates a comforting conclusion: AI becomes abundant, so software margins become enormous.
Maybe.
But this assumes the amount of inference required for each task stays roughly constant. It probably will not.
When intelligence becomes cheaper, developers do not pocket all the savings. They spend them on more intelligence.
They add another planning step. Another research agent. A critic. A verifier. A fallback model. A longer context window. Three attempts instead of one. A debate between agents because debate improves the benchmark. Continuous monitoring because the agent might miss something. Memory because every session should know what happened before.
The cost per token falls.
The number of tokens per completed outcome rises.
That is not a contradiction. It is the AI version of Jevons paradox: making a resource cheaper can increase total consumption because people find more ways to use it.
The industry is celebrating cheaper inference while designing systems that consume vastly more of it.
Stop counting agents. Count relationships.
The cost variable is not simply how many agents you deploy. It is how often they communicate, how much context they exchange and how many rounds they need before the system accepts an answer.
Five isolated agents doing five independent tasks are manageable.
Five agents constantly reviewing, correcting and updating one another are a committee.
The maths becomes ugly quickly.
With five agents, that is 10 possible relationships. With 10 agents, it is 45. With 20, it is 190.
Not every system will use every connection. Good systems should not. But once agents are allowed to delegate, debate, verify and escalate across multiple rounds, communication can move from roughly linear growth towards quadratic growth.
Then context makes it worse.
Agents rarely exchange a clean three-line instruction. They pass system prompts, task history, retrieved documents, tool outputs, intermediate reasoning, policies and previous messages. The same facts can be injected repeatedly into different model calls.
You are no longer paying only for new intelligence.
You are paying to remind the organisation what it already knows.
One request, 800,000 tokens
Consider a deliberately simple example.
A normal assistant handles a customer request with one model call and consumes 6,000 total input and output tokens.
Now turn it into an eight-agent workflow. Each agent participates in three passes. That creates 24 model calls. If each call carries an average of 25,000 tokens because it includes instructions, history, retrieved material and intermediate outputs, the system has already consumed 600,000 tokens.
Add four verification calls and a modest 15% retry rate for failed tools, weak answers or malformed output. The workflow is now near 800,000 tokens.
The customer may still receive a 600-word answer.
From the outside, the product completed one task. Inside the system, token consumption increased by more than 100 times.
The exact numbers will vary. The economic structure will not.
This is why cost per token is becoming a weak measure of agent economics.
The metric that matters is cost per verified outcome.
We are rebuilding bureaucracy in software
There is an old corporate joke that meetings are where productivity goes to die.
Agentic AI could eliminate some human meetings and then rebuild them inside the machine.
The planner asks the researcher for an update. The researcher asks the browser agent for evidence. The browser agent fails to parse a page. The researcher retries. The critic says the evidence is weak. The planner requests another search. The compliance agent objects to a sentence. The writer revises it. The evaluator gives it 0.78 when the threshold is 0.80, so the entire loop runs again.
This is bureaucracy at machine speed.
It is faster than human bureaucracy. It may still be cheaper than human labour. But cheap relative to a salaried employee is not the same as economically efficient.
If a $2 agent task could have been completed by a deterministic database query costing a fraction of a cent, the relevant comparison is not the employee. It is the software you should have written.
That distinction will separate serious operators from AI tourists.
The lazy architecture sends everything to a model. The disciplined architecture asks a harder question: where is probabilistic intelligence actually necessary?
Most business processes contain a small number of ambiguous decisions surrounded by a large amount of predictable plumbing. Models should handle the ambiguity. Conventional software should handle the plumbing.
Using an agent for every step is like hiring a management consultant to move every box in a warehouse.
The bill contains more than tokens
Token spend is only the visible layer.
An agentic workflow may also pay for search APIs, browser sessions, databases, vector retrieval, document processing, code execution, voice services, identity checks and external SaaS actions. It needs logs, traces, evaluations, security controls and human exception handling.
Then there is latency.
Parallel agents can make a system faster, but parallelism also means paying for several attempts at once. Sequential agents conserve concurrency but force the customer to wait. Providers must choose between a better user experience and a lower cost structure.
Reliability adds another multiplier. A system that is correct 90% of the time sounds impressive until a workflow contains 10 dependent steps. If each step succeeds independently 90% of the time and every step must work, the probability of a flawless run is roughly 35%.
That does not mean the workflow is doomed. Retries, guardrails and deterministic checks can improve it. But every reliability mechanism consumes more resources. The system pays a verification tax to turn probable answers into usable outcomes.
So the real equation is closer to:
Cost per outcome = inference + communication + context repetition + tools + verification + retries + human exceptions
The model call is only one line item in the synthetic payroll.
This changes the business model
The first generation of enterprise AI was sold like software: a subscription per user, per month.
That works when usage is bounded and predictable. It becomes dangerous when one user can unleash a team of agents that works continuously, calls outside tools and debates itself at 3 a.m.
In the agent economy, a seat is not a seat. It is a potential compute liability.
This creates three uncomfortable problems.
First, “unlimited agents” will become the AI equivalent of unlimited mobile data: attractive marketing with limits hidden in the fine print. Vendors offering fixed prices against variable, recursive workloads will either impose caps, degrade quality or watch their gross margins get eaten alive.
Second, outcome-based pricing will sound better than it behaves. Charging $50 for a completed task is wonderful if the system reliably spends $2. It is a disaster if hard cases trigger long research trees, repeated tool failures and human escalation. Providers will need to understand the cost distribution, not just the average. The worst 5% of tasks can destroy the economics of the other 95%.
Third, enterprise budgeting starts to look less like buying software and more like funding operations. Agent fleets consume resources in response to work volume. Their cost rises with activity, complexity and failure. Inference becomes a form of digital working capital.
That is a very different procurement conversation.
The hyperscalers should love agent chatter
For cloud and model providers, the multiplier is not necessarily a bug.
Every planner call, critique, retry and verification step is demand. The application company promises the customer an outcome. The infrastructure company gets paid for the argument that produced it.
This creates a familiar value-chain tension.
Application founders want agents to become reliable and cheap enough to support strong software margins. Infrastructure providers benefit when applications use more context, more reasoning, more modalities and more agents. Customers want predictable prices. Models remain probabilistic. Everyone wants the other party to absorb the variance.
The result will be a fight over who owns the cost-control layer.
Model providers will offer routing, caching and batch pricing. Clouds will sell agent observability and inference optimisation. Application vendors will build internal governors. Enterprises will demand budgets, approval thresholds and kill switches.
The next important category may not be another agent framework.
It may be the CFO for machines: the control plane that decides which model can be called, how much context it receives, how many retries it gets and when the system must stop thinking and return an answer.
Cheap models may win more work than smart models
The hidden multiplier also changes model competition.
If one frontier model call sits at the centre of a workflow, buyers may tolerate a premium for maximum intelligence. But if an outcome requires 40 or 100 calls, price differences compound rapidly.
That creates room for model portfolios.
A strong model plans. A cheaper model classifies. A local model extracts fields. Deterministic code calculates. A specialist model checks policy. The frontier model returns only when the workflow encounters genuine ambiguity.
The winning system will not use the smartest model everywhere. It will use the cheapest adequate intelligence at each step.
This is uncomfortable for companies whose strategy assumes every piece of cognition flows through one giant model. Agentic growth can expand the total market for inference while fragmenting where that inference runs.
It also strengthens the case for small models, on-device inference and sovereign compute. Once organisations operate millions of internal agent interactions, data location, latency and marginal cost stop being technical details. They become balance-sheet and national-infrastructure questions.
What good operators will measure
Most agent demos measure whether the system eventually completed the task.
That is not enough.
An economically serious deployment should measure:
cost per successful and verified outcome
model calls per outcome
tokens per outcome, not merely tokens per call
agent-to-agent messages per workflow
repeated-context ratio
retry and escalation rates
tool failure rates
latency at the 95th percentile
the share of steps handled by deterministic software
gross margin by task type and customer
These metrics expose the difference between useful reasoning and expensive theatre.
The best agent system may not be the one with the most agents, the longest context or the most elaborate debate. It may be the one that knows when not to think.
Use a shared state store instead of making every agent brief every other agent. Compress history. Route messages through a manager rather than creating a fully connected swarm. Cache stable facts. Put hard budgets on recursion. Use deterministic checks before model-based verification. Give cheap models narrow jobs. Escalate to expensive models only when uncertainty justifies it.
In other words, design the machine organisation the way we wish we had designed human organisations: fewer meetings, clearer authority and less duplicated work.
The hidden multiplier becomes the market
Inference will keep getting cheaper.
But the industry is moving from single answers to persistent, multi-agent systems at the same time. Unit prices are falling while the number of billable cognitive events is rising.
Both can be true.
This means the great AI cost collapse may not produce a collapse in AI spending. It may produce an explosion in consumption. Cheaper intelligence will be embedded in more processes, run more frequently and surround every important decision with planning, simulation, critique and verification.
The economic question is “How many tokens does this organisation need to reach a decision it is willing to act on?”
That is a much bigger number.
And it is why the next generation of AI winners will own the architecture, routing, budgets and controls that stop intelligent machines from spending all day talking to each other.
We spent decades trying to remove bureaucracy from companies.
We should probably avoid rebuilding it in the data centre.
Full Stack Capitalist follows AI through the entire economic stack: chips, energy, infrastructure, software, capital and power. Subscribe if you care less about benchmark theatre and more about who captures the margin.
The numerical workflow example is illustrative. Actual usage depends on model, architecture, context size, tool design and retry policy.