What is Gartner forecasting about agentic AI costs?

That agentic AI inference costs will rise more than fivefold by 2028. Gartner says routing a task to an agentic reasoning model costs at least five times a simple chatbot call, and that falling per-token prices do not offset it because agents consume far more tokens per task.

Is this an observed figure or a projection?

A projection. It is a Gartner forecast rather than measured data, and should be cited as such. The underlying mechanism — that agent loops re-send accumulated context on every turn — is something you can verify on your own usage regardless of whether the multiple holds.

Why don't falling token prices solve this?

Because the token count per task rises faster than the price per token falls. Gartner's analyst frames it as an agent needing to "constantly reason, negotiate, and question itself", each iteration carrying the accumulated context of the previous ones.

How should I measure my own agent costs?

Cost per completed task rather than cost per call or per token, and include failed and abandoned runs in the total. Excluding them is the most common way agent economics get made to look better than they are.

Gartner: Agentic AI Costs 5x by 2028

An executive in a boardroom studying a steeply rising cost curve on a tablet — DataAugust 17, 2026·2 min read·By the Automation Squad Research
Gartner: Agentic AI Costs 5x by 2028
A forecast, not a measurement. But the mechanism it describes is one you can check on your own invoice today.

By Robert MacKelfresh

Founder, Automation Squad · August 17, 2026

The short answer

Gartner forecasts that agentic AI inference costs will rise more than fivefold by 2028. It says routing tasks to agentic reasoning models costs at least five times more than simple chatbot calls, and that falling per-token prices do not compensate because agents consume far more tokens per task.

Cost model

Work out your cost per completed task, not per token

The forecast is Gartner's. The arithmetic below is yours, and it either supports the claim or it doesn't — which is more useful than the forecast either way.

Pick one workflow you have running both ways
Something you do with a single call and something you do with an agent loop. If you have no agentic workflow yet, this exercise costs you nothing and tells you what to expect.
Measure tokens per COMPLETED task, not per call
This is where the cost hides. An agent that takes eleven calls to finish is eleven calls of input, and each one carries the accumulated context of the last. Per-call figures make that invisible.
Include the runs that failed
Cost per completed task means dividing total spend by tasks that actually finished. Abandoned runs and retries are part of the cost of the ones that worked, and excluding them is the single most common way these numbers get flattered.
Compare against the human baseline honestly
Five times the inference cost is irrelevant if the task previously took a person forty minutes. It is decisive if the task was already automated another way. The multiple only means something next to what it replaced.
Re-run it monthly, because this is the number that drifts
Agent context grows, prompts accumulate, retries creep in. A cost-per-task figure from three months ago is not your cost per task now.

Gartner's claims	Status
Agentic inference costs rise 5×+ by 2028	Forecast
Agentic reasoning costs ≥5× a simple chatbot call	Forecast
Falling per-token prices do not compensate	Forecast
40% of organisations will demote or decommission AI agents	Earlier Gartner forecast
At least half of genAI projects will overrun budgets	Earlier Gartner forecast

Take it with you

AGENTIC AI COST — work out YOUR number
Gartner forecast, Aug 17 2026
Source: https://automationsquad.com/news/gartner-agentic-cost-forecast/

WHAT GARTNER SAYS (all FORECAST, not measurement)
  Agentic inference costs rise 5x+ by 2028
  Agentic reasoning >= 5x a simple chatbot call
  Falling per-token prices do NOT compensate — agents use far
  more tokens per task
  (earlier Gartner forecasts: 40% of orgs will demote or decommission
   AI agents; at least half of genAI projects overrun budgets)

THE MECHANISM — worth checking on your own invoice
An agent "must constantly reason, negotiate, and question itself."
Every loop re-sends accumulated context.

YOUR ARITHMETIC
[ ] Pick one workflow. Single call vs agent loop.
[ ] Measure tokens per COMPLETED TASK, not per call
      calls to finish: ____   avg tokens/call: ____
[ ] INCLUDE FAILED RUNS in the total
      total spend / tasks that actually finished = ________
      (excluding failures is the #1 way this gets flattered)
[ ] Compare to what it replaced:
      previously a human?  -> 5x inference may be irrelevant
      previously automated? -> 5x is decisive
[ ] Re-run monthly. Context grows, retries creep in — this number
    drifts more than any other cost you track.

Gartner published a forecast on August 17, 2026 that agentic AI inference costs will rise more than fivefold by 2028.

The facts: Gartner says routing tasks to agentic reasoning models costs at least five times more than simple chatbot calls, and that declining per-token prices do not compensate, because agents consume far more tokens to complete the same work. Gartner senior director analyst Will Sommer frames the mechanism as an agent needing to "constantly reason, negotiate, and question itself." The firm has previously forecast that 40% of organisations would demote or decommission AI agents, and that at least half of generative AI projects would overrun their budgets. All of these are projections rather than measurements.

Automation Squad's take: treat the number as a hypothesis and the mechanism as the useful part. Analyst forecasts about 2028 are worth little on their own — the value here is that it names precisely why agent economics surprise people, which is that the unit everyone quotes and the unit that lands on the invoice are different. Per-token prices have genuinely fallen and continue to. Tokens per completed task have risen faster, because every loop re-sends the accumulated context of the previous one, and a workflow that needs eleven turns pays for the conversation eleven times. There is a neat coincidence in the timing worth noticing: the same week this forecast landed, OpenRouter shipped a dashboard that measures exactly this. The problem and a tool for looking at it arrived within twenty-four hours of each other.

Run this now: take one workflow you run as an agent loop and calculate cost per completed task — total spend divided by the number of tasks that actually finished, with failed and abandoned runs left in the numerator. That last part is where most people quietly flatter themselves. Then compare it against whatever the task cost before. If it replaced forty minutes of a person's time, five times the inference cost may be irrelevant; if it replaced something already automated, the same multiple is decisive. Set a reminder to redo it in a month, because this is the number that drifts most.

Questions people are asking

What is Gartner forecasting about agentic AI costs?: That agentic AI inference costs will rise more than fivefold by 2028. Gartner says routing a task to an agentic reasoning model costs at least five times a simple chatbot call, and that falling per-token prices do not offset it because agents consume far more tokens per task.
Is this an observed figure or a projection?: A projection. It is a Gartner forecast rather than measured data, and should be cited as such. The underlying mechanism — that agent loops re-send accumulated context on every turn — is something you can verify on your own usage regardless of whether the multiple holds.
Why don't falling token prices solve this?: Because the token count per task rises faster than the price per token falls. Gartner's analyst frames it as an agent needing to "constantly reason, negotiate, and question itself", each iteration carrying the accumulated context of the previous ones.
How should I measure my own agent costs?: Cost per completed task rather than cost per call or per token, and include failed and abandoned runs in the total. Excluding them is the most common way agent economics get made to look better than they are.