Agent Workflow Costs Rise as Token Prices Fall

An AI application lead puts two bills side by side at the quarterly budget review. On the left is the 2024 model usage record; on the right, last month. The traffic is the same question-and-answer work her team ran two years ago, and the unit price has dropped 280-fold. What she has to explain is not that number. It is the total on the right, which is higher than it was two years ago.

The unit price fell 280-fold. The bill did not follow

The price decline is well documented. Mirantis relays a Deloitte 2026 analysis of AI infrastructure: inference costs fell 280-fold over two years. Over the same window the overall AI bill kept climbing, because agentic systems generate continuous inference demand, and that demand grew faster than the price fell.

Zylos put the same movement into numbers in its April 2026 industry review. GPT-4-class capability cost $30 per million tokens in March 2023. Today it costs about $0.40. On unit price, the bill should be smaller.

On the total, it is not. This means the assumption we carried through the last two years, that models keep getting cheaper and so AI keeps getting cheaper, holds at the token layer and fails at the bill layer.

The lead stares at both lines for a while. The unit-price curve points down. The total curve points up. Then she notices the two bills are not measuring the same thing. The left one counts answers. The right one counts tasks.

Open the two bills and the gap is visible on the page. She ran the same workload in both years, and the per-call rate had collapsed. She checked the pooled total and it had gone up, not down. She wrote the two numbers on one line and asked which one the board would see first. Nearly every team answers the same way, and that answer is the whole problem.

The billing unit moved from an answer to a completed task

The change is in the unit of account. In Zylos’s review, a single chatbot call costs about $0.001. A multi-step agent task that plans, retrieves context, invokes tools, reflects on its output, and corrects itself costs $0.10 to $1.00 per completion. Same model family, different unit: 100x to 1,000x.

Gartner added a supporting figure in March 2026: agentic tasks consume 5x to 30x more tokens per task than a standard chatbot. On August 17, Gartner moved the trend onto the finance calendar. Inference costs per agentic workflow will rise more than fivefold through 2028.

Put the two together and the opening bill explains itself. The price falls at the token layer; the bill rises at the task layer. You buy tokens, you consume completed tasks, and the conversion between the two widens every year.

So the real question in a budget review is task volume. One answer runs once. One task runs a dozen tool calls. Answer volume scales with people; task volume scales with process steps and retries. Change the unit of account and the ledger changes with it.

An answer finishes in one round. A task runs a dozen tool-call rounds, and it may retry, wait on an external API, or re-run earlier steps after one bad judgment. Count all of that and you see what a task really costs. What we pay for is not the call. It is the completed task.

Where the money goes inside one completed task

To bring task cost down, you first need to know where it goes. Four variables decide it: step count, the model tier at each step, retry rate, and the context carried into every round.

Only one of these four is negotiable in procurement: the unit price. The other three live inside our own architecture. So the main battleground for the bill is not the negotiating table. It is the runtime.

What one bill actually looks like

Take the opening bill apart. Say a task runs twelve tool calls, and each round carries an average of 20,000 tokens of context, most of it repeated between rounds. Price it at today’s roughly $0.40 per million tokens.

One run of that task spends close to a tenth of a dollar on the input side alone. If one or two calls fail and re-run, that tenth is paid again. Per call it looks negligible. Per completed task it is a line item.

Now multiply by volume. A department running 2,000 tasks a day runs about 60,000 tasks a month. No downward price curve outruns that. Which means the real control points are the round count per task and the context carried into each round.

That is why the first cut should land on context and rounds, not on the unit price. Cutting the unit price is the easiest thing to do and usually the least effective thing to do.

Go a layer deeper. Of twelve rounds, suppose ten repeat 80% of the history. About 20% of the tokens carry new information; the other 80% are paid for more than once. Compressing that saves more than a quarter of price negotiation, which is why an engineering afternoon beats a procurement debate on this bill.

Three numbers to compute first

If you have one afternoon, compute three numbers: average rounds per task, the share of rounds that need a frontier model, and the retry rate. You do not need precision. The order of magnitude is enough.

Open your own traces before you trust any benchmark. We pulled one task type, counted the rounds by hand, and saw retries we had never budgeted for. We checked the context on each round and found most of it repeated. We tried the same task on a cheap tier and the bill moved further than any discount had. Which means the useful test is your own workload, not the vendor page.

Take a customer-service ticket task. It averages 8 rounds, of which 1 needs frontier judgment and the other 7 are extraction and formatting. The retry rate is about 5%, so roughly one task in twenty re-runs. Those numbers define the shape of the task cost and where routing can save.

Move those 7 rounds from the frontier tier to the cheap tier and the bill drops visibly. Negotiate a 10% discount on the unit price and the bill barely moves. For the same engineering afternoon, the first option returns dozens of times more.

Look at the 2026 price band. In the Digital Applied review, the cheapest usable model runs about $0.44 per million tokens, and the most capable runs $30 for input. The same prompt lands two orders of magnitude apart depending on the tier. That is the case for a routing layer, and the reason we price by task instead of by token.

Why this decline did not arrive like the others

Over the past decade, enterprises have absorbed several unit-cost collapses. Cloud compute got cheaper, storage got cheaper, bandwidth got cheaper, and every one of those declines arrived as a lower bill. It is natural to assume AI follows the same path.

What is different this time is the shape of usage. Cloud usage scales with the business: one user maps to a fairly fixed amount of storage and bandwidth. Agent usage scales with tasks and rounds. Take a workflow from three steps to ten and token consumption does not grow by the same order. It can multiply.

In other words, the fall in unit price is being eaten by the rise in task count, and the meal is not finished. That is the arithmetic in the opening bill. Stare at the unit price and you will keep getting the right answer to the wrong question.

There is another difference worth noting. Cloud cost curves are smooth: demand rises, the bill follows. Agent cost curves are stepped: add two steps to a workflow and the bill can double. The same task, before and after a redesign, can differ several times over.

A pilot ledger and a production ledger are two kinds of math

A pilot ledger is easy to balance. One demo task, run fifty times, with a unit price that fell 280-fold, is cheap no matter how you compute it. At scale the ledger switches to different math: fifty tasks times a thousand runs each, with retries, tool calls, and extra judgment rounds in between.

IDC’s FutureScape 2026 report, relayed by Zylos, sets an unflattering baseline. Even organizations with 1,000+ employees and a dedicated FinOps function will still underestimate AI infrastructure costs by up to 30%. Over the same period, the average enterprise AI budget rose from $1.2M a year in 2024 to $7M in 2026.

Put those two numbers together and the conclusion is plain. The fall in unit price did not pull total cost down, and the rise in budget says so. What we saved is the cost of each token. What we spent more on is the cost of each task, and task count grew faster.

The payoff is real, but it is not in the demo. Mirantis relays McKinsey’s April 2026 analysis: agentic AI can automate 60% to 80% of routine infrastructure work and cut run-rate cost by 20% to 40%. That number lives in production, in the monthly bill, not on the day of the demo.

So a cheap demo is not a cheap rollout. Between the two sit task volume, retry rate, and tool-call count. Multiply them in and the pilot return usually falls sharply. Optimize them out and the return comes back.

Our judgment: unit price is a procurement metric, cost per task is an operating metric

Our judgment is this. The token unit price is a procurement metric. What one completed task costs is an operating metric. The difference between them is who answers for the outcome. A procurement metric can be pressed down at the table. An operating metric can only be built inside your own architecture.

A team that only squeezes the unit price will be eaten by its own bill as volume grows. The price falls, the task count rises faster, and the product of the two still rises. The problem is in the unit of account, not in the model’s list price.

The right move is to route by task tier. Give most steps to a cheap tier and escalate to a frontier model only at the judgment points. In the Digital Applied engineering review from June 2026, teams that deployed a routing layer report bill reductions of 40% to 85%. In RouteLLM’s controlled evaluation, an 85% cost cut preserved 95% of GPT-4 quality. Those are reported and benchmark figures, not guarantees, but the direction is clear.

We have tested this judgment in our own operation. OOMeta’s self-built agent runtime is layered: most steps run on a cheap tier, and only the judgment points escalate to a frontier model. Layering does not change task count. It changes the average cost per task and the slope at which the bill rises with volume.

Where this judgment stops

Our judgment has edges. Task-tier routing fits workflows with many steps and few judgment points, such as data cleanup, document processing, and routine operations. For tasks where every step needs strong reasoning, such as complex code refactoring or high-stakes decisions, there is little room to layer, and the bill will rise in step with usage.

So decide which kind of workflow you have first. If most steps are deterministic, routing saves a large share. If every step is a judgment call, the question is whether the task belongs on an agent at all, not which model tier to use.

Move the budget yardstick to the completed task

What changes is a unit of account. Four steps, in this order.

Order matters. Build per-task metering first, then let the routing layer execute the tiering, and only then return to the unit-price negotiation. Reverse the order and you will press the price down without data, then meet the same bill again in your first quarter at scale.

Who owns the cost of each task

A new unit of account still needs an owner. We suggest one owner per task type, the way you assign an owner to each line of cloud spend. A task with no owner drifts, and it drifts until nobody will claim it.

The work is not complicated. List your tasks, sort by call volume, and note that the top 20% usually carry 80% of the spend. Put the owner, average rounds, retry rate, and cost per task on one sheet, and read it once a month. It does not have to be precise to show a trend.

Too many teams leave this to quarter-end. By then the bill is out and the only thing left to change is the next quarter. If you do one thing today, build the cost-per-task column, even while its numbers are still rough.

The next time you open the model bill, the first question is not whether the unit price fell. It is what a completed task cost. If your team still cannot answer that, the bill does not have a real owner yet.

References

• Gartner: AI Inference Costs Per Agentic Workflow Will Increase More Than Fivefold Through 2028 (2026-08-17) — gartner.com/en/newsroom
• Mirantis: Understanding Agentic AI Infrastructure (2026-06-09, relaying Deloitte and McKinsey 2026 analyses) — www.mirantis.com/blog/agentic-ai-infrastructure
• Zylos Research: Inference Economics: AI Agent Compute Markets in 2026 (2026-04-13, relaying IDC FutureScape 2026) — zylos.ai/research/2026-04-13
• Digital Applied: LLM Model Routing in 2026 (2026-06-14) — www.digitalapplied.com/blog/llm-model-routing-2026