A single question to a language model costs a fraction of a cent. An agent that books a travel itinerary, searching flights, checking seat availability, reading cancellation terms, and confirming a selection, makes dozens of model calls, each one re-reading everything that came before. That accumulation is what separates a cheap API query from a real line item.
Token pricing sets the floor
Every commercial LLM API bills by token, the subword units a model uses to process text. Input and output are priced separately; output costs more because generating each token requires a full forward pass through the model, while reading input is cheaper. For detail on how this works, see what tokens are and how they’re counted.
As an illustrative example, as of mid-2026 a fast small-model tier might run around $1 per million input tokens while a frontier reasoning tier might reach $50 per million output tokens; prices vary by provider and change over time. Fast, capable models for routine tasks occupy the low end; frontier reasoning models occupy the high end.
The loop multiplies that floor
An agent doesn’t make one call. It reasons, invokes a tool, reads the result, reasons again, and continues until the task is complete or a constraint stops it. What most cost estimates miss: each new reasoning step sends the full conversation history as input. Step 15 in a loop includes steps 1 through 14.
Research on tool-calling chains in LLM agents found that context amplification through chained tool calls can increase per-query costs by hundreds of times compared to a single-turn query (Zhou et al., arXiv:2601.10955, 2026). A 20-step agent run is not 20 times the cost of a single call. It costs substantially more, because each step’s input grows as the run continues.
Tool calls add external charges
When an agent calls a web search API or triggers an external service, those calls carry separate costs beyond token pricing. Some services charge per request; others charge by data volume. An agent that makes 10 tool calls per task pays 10 additional charges, plus whatever model inference the tool results trigger in the next reasoning step.
Subscription products absorb but constrain usage
Not all AI agent access is billed by token. Many commercial products charge a flat per-seat or per-month fee and absorb API costs within that price, typically enforcing usage caps or rate limits in place of per-token charges. Whether a subscription is cheaper depends on task frequency: light, occasional use often favours subscriptions; high-volume, long-running tasks often favour usage-based access at scale.
Task scope is the real variable
Token pricing and loop mechanics are predictable. What isn’t is how many steps a specific task actually requires. An agent given a bounded task like “extract the deadline from this contract” completes it in a handful of steps. An agent tasked with “review this project and flag everything that needs attention” may run far longer, depending on what it encounters.
This is the non-obvious cost driver: scope discovered during execution, not scope declared at the start. Open-ended tasks and large attached documents drive bills more than the nominal cost-per-call implies. Matching the model to the actual capability a task requires, rather than defaulting to the most powerful available, is the most direct lever on what agents cost at scale.
Sources
- Zhou, K. et al. Beyond Max Tokens: Stealthy Resource Amplification via Tool Calling Chains in LLM Agents. arXiv:2601.10955, 2026.
Questions, answered
Why are AI agent costs hard to predict?
An agent's cost depends on how many reasoning steps it takes to complete a task, which only becomes clear during execution. A task you expect to resolve in five steps may take fifteen if the agent encounters unexpected complexity, and each additional step re-sends the full conversation history as input.
Is token pricing the same for every AI provider?
No. Rates vary significantly by provider and model tier. As an illustrative example, as of mid-2026 a fast small-model tier might run around $1 per million input tokens while a frontier reasoning tier might reach $50 per million output tokens, with output tokens typically priced three to five times higher than input.
Do subscription-based AI tools hide token costs?
Often yes. Per-seat or per-month products absorb API costs and may enforce usage limits or rate caps instead of charging by the token. Whether that is cheaper depends on task intensity: light, occasional use tends to favour subscriptions; high-volume, long-running tasks often favour usage-based pricing.
What is the single biggest lever on AI agent cost?
Model choice for each step. Running every reasoning step through a frontier model costs far more than routing simpler steps to a smaller, cheaper model. Most agent frameworks allow per-step model selection, making this the most direct cost control available.
Brief is a team of AI associates you direct in plain language. Opening to a small group at a time.
Request access