AI agent costs: why agents burn tokens, and how to cut the bill
AI agents are the single biggest driver of token spend in 2026. The same task that costs a fraction of a cent as a one-shot chatbot reply can cost 30× more as an agent. Here's exactly why, and how to bring it down.
Why agents cost so much more than chatbots
A chatbot sends one prompt and gets one reply. An agent runs a loop: call the model → decide an action → call a tool → feed the result back → call the model again, repeating until the task is done.
The expensive part is that the agent re-sends the entire accumulated context, the system prompt and instructions, plus the full history of every step, to the model on every iteration. By step 20 of a task, you are paying for the same early context twenty times over. Re-sent context is, by some 2026 audits, the majority of an agent's bill.
A worked example
Single-shot vs 5-step agent
A task that takes 3,000 tokens as one model call can use 12,000–18,000 tokens as a 5-step agent loop, because each step re-sends the growing context.
A mid-complexity support agent might use 2,000–5,000 tokens per conversation. At roughly $3 input / $15 output per million tokens, that's about $0.01–$0.03 each, which sounds tiny until you run 100,000 conversations a month and it becomes $1,000–$3,000/month.
Plug your own per-step token counts and call volume into the cost calculator to estimate an agent workflow: set input/output tokens to one loop iteration, and set "requests/day" to (loops per task × tasks per day).
How to cut agent costs (50–70% is realistic)
2026 audits of teams running agents in production consistently point to four levers:
- Route by difficulty (tiered models). Send easy steps, classification, extraction, routing, to a cheap model, and reserve a frontier model for the genuinely hard reasoning. Analyses of billions of enterprise calls found a blended cost near $2.31 / 1M tokens for tiered setups versus $18.40 / 1M for routing everything to frontier models, an ~87% gap. Compare cheap vs frontier prices →
- Use prompt caching. Since agents re-send the same system prompt and context every step, caching is tailor-made for them. Cached input is billed at roughly 10–25% of the normal rate and can cut agent bills 50–70%.
- Trim the re-sent context. Summarize old steps, drop irrelevant history, and keep the working context lean instead of appending everything forever.
- Batch what isn't urgent. For non-interactive agent jobs, a batch API often costs around half the standard rate.
The takeaway
Agent cost is mostly an architecture decision, not a price you're stuck with: how many loops you run, how much context you re-send each loop, which model handles each step, and if you cache. Pick the right model per step and cache aggressively, and the same agent can cost a fraction of the naive version.
Cost calculator Compare models Cheapest models Cut API costs
Figures are approximate ranges drawn from 2026 industry analyses and reports, shown for orientation only, not exact quotes, and they change frequently with model prices and your own workload. Always confirm current per-token prices on each provider's official page and estimate your own usage with the live TokenSwarm calculator. This guide covers agent costs; it is not a tutorial on building agents. Not affiliated with any provider.