How Much to Charge Clients for an AI Agent: Start From Your Cost Floor

Updated 2026-10-25. Checked against official provider docs and pricing pages on 2026-10-25.

There's no honest single number for what to charge. It depends on what the agent is worth to the client, your market, and your build effort. This guide doesn't give you "market rates." It gives you something more useful: a method that guarantees you never price an agent below what it costs you to run.

The method has three parts:

1. Cost floor: what the model API costs you per client per month.
2. Margin: what you add on top for your time, your risk, and your profit.
3. Overage caps: what protects you when one client uses far more than you planned.

Value-based pricing (charging for the outcome) sits on top of this. The floor just stops you from losing money on a busy client.

Step 1: Compute your cost floor

Agent API cost follows one formula:

Monthly API cost = tasks per month × model calls per task × (input tokens × input price + output tokens × output price) ÷ 1,000,000

You need four usage numbers. Get them from a pilot or your logs, not from a guess:

• Tasks per month: conversations, tickets, or jobs handled for this client.
• Model calls per task: agent loops call the model several times (planning, tool calls, final answer).
• Input tokens per call: system prompt, tool definitions, history, and tool results.
• Output tokens per call: including reasoning tokens, which are billed as output.

Worked example (example assumptions, not benchmarks)

Say a client's support agent handles 1,500 tasks a month at 8 model calls per task, with 3,000 input and 400 output tokens per call. These are illustrative numbers; swap in your own.

Prices per 1M tokens are from TokenTab's models.json, retrieved 2026-10-25.

*Google promotional prices. Gemini 4 Argon's standard price rises to $4/$20 on Dec 1, 2026, which doubles its row. Gemini 3.8 Flash's price runs through Dec 31, 2026 and rises to $1.50/$7.50 on Jan 1, 2027, which doubles its row too.

†DeepSeek Flash is shown at its peak rate. DeepSeek's off-peak rates are half (Flash: $0.15/$0.60); peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday.

Two things stand out. First, model choice moves the floor 100x for the same workload. Second, this is list-price, uncached cost. Prompt caching and batch processing (for non-urgent work) can lower it, but only count those savings once you've measured them.

To run this with your own numbers in a few seconds, plug them into the TokenTab AI agent cost calculator. It shows per-task, per-day, and per-month cost across 32 models.

Add everything else that scales with the client

The model API is rarely your only per-client cost. Add your other variable costs at their actual billed rates: hosting, a vector database, telephony, speech-to-text and text-to-speech for voice agents, search APIs, and logging. Then:

Cost floor = model API cost + other variable costs + any per-client fixed costs

Before you set a price, measure the cost side first with how to estimate AI agent costs.

Your margin improves if a stable system prompt is cached; see does prompt caching save money for AI agents?

Step 2: Plan for the busy month, not the average one

Usage is lumpy. A client's busiest month, a launch, or an agent stuck retrying a broken tool can multiply your cost. Price from a high-usage estimate, not the mean. If you have logs, use a high percentile of monthly usage. If you don't, pick a stated buffer.

In the example, if the busy month is 2x the baseline (an example multiplier), the Sonnet 5.5 floor becomes $240, while Haiku 5.5 or GPT-6 Luna becomes $12.

Step 3: Add margin

Margin is a business decision. This guide won't tell you what's "normal." Use the formula to turn whatever margin you choose into a price:

Price = cost floor ÷ (1 − target gross margin)

For example, with an example target margin of 60% on a $240 floor: $240 ÷ 0.4 = $600 a month. Change the margin and the number changes. The point is that the price is derived from your cost, so you can defend it.

Your margin has to cover the costs that don't show up on the API bill: monitoring, prompt maintenance, model migrations when a provider retires a model, support, and your time.

Step 4: Put overage caps in the contract

The single biggest pricing risk with agents is open-ended usage on a flat fee. Common ways to structure protection:

• Included allowance: the fee covers N tasks, conversations, or minutes a month.
• Overage rate: a per-unit price above the allowance, derived from your per-task cost plus margin.
• Hard cap or approval gate: usage above a set point needs client sign-off.
• Pass-through: the client pays the provider directly, or you bill API usage at cost plus a stated markup.
• Model clause: state which model tier the price assumes, because switching from a small model to a frontier one can multiply cost.

On the technical side, back the contract with provider spend limits per client (a separate project, workspace, or key for each) and per-run budgets in code, so a runaway loop can't eat a month's margin. Our guide on stopping runaway agent loops covers both.

Step 5: Re-check when prices change

Model prices change, and promos expire. Gemini 4 Argon's price doubles on Dec 1, 2026, and Gemini 3.8 Flash's doubles on Jan 1, 2027. Recompute your floor whenever a provider changes prices, and write a price-adjustment clause into longer contracts.

A one-page pricing checklist

• Measured tasks per month, calls per task, and tokens per call from a pilot
• Monthly API cost per model, at list prices
• Other variable costs added at their real rates
• A high-usage multiplier applied
• Margin chosen; price = floor ÷ (1 − margin)
• Allowance, overage rate, and cap written into the contract
• Separate API key or project with a spend limit per client
• Review date set for price changes

Sources

Retrieved 2026-10-25.

• TokenTab models.json (lastUpdated 2026-10-25): /models.json
• Google Gemini API pricing (promo end dates): https://ai.google.dev/gemini-api/docs/pricing
• OpenAI pricing: https://developers.openai.com/api/docs/pricing
• Anthropic pricing: https://platform.claude.com/docs/en/about-claude/pricing
• OpenAI spend limits: https://developers.openai.com/api/docs/guides/spend-limits
• DeepSeek pricing (peak and off-peak): https://api-docs.deepseek.com/quick_start/pricing
• Anthropic spend limits and workspaces: https://platform.claude.com/docs/en/api/rate-limits