Token prices keep falling. Unit prices are expected to drop 30–50% over 2027. So why are AI bills going up?
Because usage is growing faster than prices are falling, and more and more of that usage is agentic. We built a 2027 Token Budget Estimator to model this. The results surprised even us.
The math most budgets use
Most AI budgets look like this: requests × average tokens × list price. That works for chat and RAG, where one prompt gets one answer.
It doesn’t work for agents.
Take a company running 1 million AI requests in January 2027, growing 8% a month. With our default assumptions, its bill goes from $121K in January to $282K in December, or $2.3M for the year including a 20% contingency. Prices fell. The monthly bill more than doubled.
The “average” request costs $0.10, but that hides a 60x spread. A document-extraction request costs well under a cent. A coding request costs $0.39.
The quadratic tax
Here’s why. An agent doesn’t send one prompt. It takes a step, calls a tool, reads the result and takes another step, re-sending the entire context each time. Input tokens grow with the square of the step count.
A simple customer support agent (look up the order, check the policy, issue the refund) starts with about 1,000 tokens of context. Over five steps it consumes roughly 38,500 input tokens.
A coding agent is worse. With an 8,000-token starting context, 15 steps and some retries, one task consumes about 362,000 input tokens. Ask the same question as a single prompt and it’s 4,000.
Small changes compound:
- 10% more retries means about 17% more tokens.
- Twice the steps means about 3.3x the tokens.
That’s the quadratic tax, and most budgets don’t model it.
Where the money actually goes
In our model, agentic tasks are 33% of requests but 93% of spend.
Software engineering alone is 20% of requests and 78% of spend, because coding agents run long loops on frontier models. Document processing, content, back-office automation and customer support together are 65% of requests and under 10% of the budget.
The lesson: average cost per request is a misleading number. Budget by task area, and by how agentic each one is.
Levers before routing
Some savings are available to everyone:
- Prompt caching. Agent loops re-send the same context every step, so cache hit rates run 50–80%. In our model, caching cuts agentic input cost by 60%.
- Batch pricing. Move work that isn’t real-time to async batch APIs.
- Hosting choices. Enterprise hosting and compliance typically add 15–20%. Know what you’re paying for.
These help. But the biggest lever is which model handles each request.
Routing: one rule doesn’t fit every task
Most teams already mix models. In our example, about 38% of requests go to frontier models and the rest to mid-tier, open-source and small models.
The obvious next step is a rules-based gateway that sends a fixed share of traffic to small language models (SLMs). The problem is that it can’t tell tasks apart, so it has to be cautious everywhere. In our model, a gateway moving a flat 15% of frontier traffic to SLMs still costs more than the company’s existing hand-tuned mix.
Task-based routing works differently. It decides, task area by task area, how much work a small model can handle: 90% of document extraction, 85% of customer support, but only 60% of coding. Across the workload, that moves 78% of requests to SLMs.
For our example company:
- **Current model mix:** $1.91M a year
- **Task-based routing:** $0.94M a year
- Net savings: $873K, or 46%, after an assumed $100K platform cost
Against an all-frontier baseline the savings reach $1.65M, but the comparison that matters is against what you spend today.
One caveat, stated plainly: these numbers assume small models meet the quality bar for the work moved to them. That has to be validated task by task, and it’s exactly why one-size-fits-all routing stays conservative.
Build your own budget
Every number above comes from illustrative defaults, and yours will differ. The estimator has three tabs (Assumptions, Budget and Routed), and every input is a yellow cell. Start with these five:
- Monthly request volume and growth rate
- Your current model mix by task area
- Agent step counts and context sizes
- How much of each task area could run on smaller models
- Your contracted rates
Download the 2027 Token Budget Estimator
Summary
Cheaper tokens won’t save your 2027 budget. Agents multiply token usage faster than prices fall, and the cost concentrates in a few task areas. The companies that control inference spend will be the ones that design their workloads deliberately: shorter loops, better caching and the right model for each task.