What does an AI agent actually cost to run? The honest answer has a formula rather than a number: cost per task = (input tokens + output tokens) × model price × steps, summed over every model call the task triggers — and the steps term is where estimates go wrong, because agents loop, retry, delegate, and re-read their own context. This post puts current public numbers into that formula, works through representative scenarios, and turns the results into budget-setting heuristics.
Two ground rules. First, all model prices below are providers' public list prices as of this writing (July 2026) — prices in this market change frequently and usually downward, so verify against provider price pages before committing budgets; the numbers here are for reasoning about magnitudes, not procurement. Second, benchmarks inform budgets but do not enforce them — for that, see how to set budgets for AI agents, and estimate your own workload mix with the AI cost calculator.
The 2026 price landscape
The market has settled into visible tiers. Representative published list prices (per million tokens, input / output):
| Tier | Representative models | Input | Output |
|---|---|---|---|
| Budget / high-volume | GPT mini class, Gemini 2.5 Flash, Claude Haiku class | ~$0.25–$1 | ~$2–$5 |
| Mid / workhorse | Gemini 2.5 Pro, Claude Sonnet 5, GPT-5.4 | ~$1.25–$2.50 | ~$10–$15 |
| Premium reasoning | Claude Opus class | ~$5 | ~$25 |
| Top-end | Claude Fable 5 / Mythos 5 | ~$10 | ~$50 |
Tier footnotes from the providers' own price pages: Claude Sonnet 5 carries introductory pricing ($2/$10) through August 31, 2026, then $3/$15; Gemini 2.5 Pro's $1.25/$10 applies to prompts up to 200k tokens; GPT-5.4 lists at $2.50/$15.
Three structural facts matter more than any single figure. The budget-to-top-end spread is roughly 25–40x ($0.25 vs $10 on input, $2 vs $50 on output) — which makes routing tasks to the cheapest sufficient model the largest cost lever available, ahead of any prompt optimization. Output costs 4–8x input across vendors — agents that generate verbosely pay for it directly, and step-limiting an agent's "thinking out loud" is a real economy. Input dominates volume in agentic workloads — every step re-sends conversation history, retrieved documents, and tool schemas, so context discipline (summarization, selective retrieval, prompt-cache use where providers offer it) compounds across every step of every task.
Worked scenarios
Applying the formula with mid-2026 list prices — your mileage will vary with prompt design, but the magnitudes are robust:
Simple support-agent task (classify + draft a reply; 1–2 model calls; ~3k input, ~500 output tokens on a budget-tier model): roughly $0.002–$0.005 per task. At 10,000 tasks/month: $20–$50/month. The headline: simple, well-scoped agents are nearly free — until step counts drift.
Multi-step research/summarization task (5–10 steps with retrieval; ~40k cumulative input, ~4k output on a mid-tier model): roughly $0.09–$0.16 per task. At 5,000 tasks/month: $450–$800/month. Context accumulation, not output, is the cost center.
Agentic coding task (10–30 steps, large file context, mid-to-premium-tier model; 100k–500k cumulative input, 10k–30k output): roughly $0.40–$3.00+ per task, with heavy tasks and top-end models several times that. A team of agents doing hundreds of such tasks daily is a four-to-five-figure monthly line item — real money, worth real routing decisions.
The pathological case (retry loop: 50k tokens/iteration, one iteration/minute): roughly $15–$30 per hour per stuck loop on premium tiers, and up to ~$45 on a top-end model — small enough to hide in a big bill, large enough to matter multiplied by fleet size and days-to-detection. This is why the runaway-spend threat model treats missing caps, not exotic attacks, as the root condition.
From benchmarks to budgets
The benchmarks' practical use is calibrating enforcement:
- Set per-agent caps from measured per-task cost. Cap at expected monthly volume × per-task cost × a 2–3x headroom factor. Tighter than that fights normal variance; looser than that stops nothing.
- Layer scopes. Per-agent caps as the first line, workflow and team caps above, an organization ceiling as the backstop — the architecture in budgets and quotas, with the most restrictive policy winning at dispatch.
- Enforce by reservation, not alert. Post-hoc alerts price in days of drift; reservation-based enforcement blocks the task before the tokens burn. Pair caps with rate limits for the fast-loop case — the two levers are complementary, per budgets vs rate limits.
- Attribute before you optimize. Per-agent, per-workflow cost attribution tells you which of the scenarios above you are actually running; optimization without attribution is guesswork. The discipline is covered in FinOps for AI agents.
- Re-benchmark on change. Model swaps, prompt rewrites, and context-window changes move per-task cost by multiples, not percents. Treat the benchmark as a living number — and rerun your assumptions through the cost calculator when they change.
What the benchmarks do not capture
List-price arithmetic understates total cost in known ways: tool-call and downstream API charges (search, data providers, execution sandboxes) ride alongside token spend; provider-side features (caching discounts, batch tiers) can cut effective prices substantially for eligible workloads; and the largest omission is unattributed spend — agents outside governance whose costs surface only on the invoice. Cost governance and agent inventory are the same project viewed from two directions.
Common questions
How much does a typical AI agent cost per month? There is no typical agent, but the bands are: simple high-volume task agents in the tens of dollars per month; multi-step research/processing agents in the hundreds; heavy agentic-coding fleets in the thousands to tens of thousands. Per-task cost times realistic volume — with a 2–3x variance factor — is the honest estimate, and the AI cost calculator does that arithmetic for your mix.
Which single decision most affects agent cost? Model routing: matching each task class to the cheapest sufficient model tier. With a roughly 25–40x list-price spread between budget and top-end tiers as of mid-2026, sending classification tasks to a premium reasoning model is the most expensive habit an agent estate can have.
Why do agent costs so often exceed estimates? The steps term: estimates model the happy path, while production agents retry, loop, expand context, and delegate. Every step re-sends accumulated context, so cost grows superlinearly with step count. Caps and rate limits exist to bound exactly this gap between modeled and actual behavior.
Are these prices going to be accurate next quarter? Assume not — this market reprices frequently, usually downward, and new tiers appear several times a year. The magnitudes and ratios (output over input, tier spreads, context dominance) have been far more stable than absolute figures. Verify list prices at decision time; keep the formula.