How much does an AI agent cost per month? (Oct 2026, by agent type)

"An AI agent" can mean a coding agent that reads a whole repo or a support bot that answers in three short turns, so the monthly bill can range from a few dollars to thousands. Below are four common agent types. Each is built from published workload numbers (tokens and LLM calls per task), priced with the TokenTab calculator formula:

Cost per task = calls × (input tokens × input price + output tokens × output price) / 1,000,000. Monthly = per task × tasks per day × 30.

About volumes: unless a source gives one, the Light (10/day), Medium (100/day) and Heavy (1,000/day) task volumes are example assumptions, not measured data. Swap in your own in the calculator. Prices shown are input / output in $ per 1M tokens.

1. Coding agent

SWE-Effi averaged agents on SWE-bench Verified, across resolved and unresolved issues:

• OpenHands: 335K–897K input tokens, 1.7K–16.7K output tokens and 19.5–34.1 LLM calls per task.
• SWE-agent: 194K–8.1M input tokens and 29–181 LLM calls per task.

These are per-task totals, so the calls are already counted. The table prices the low end (335K in, 1.7K out) and the high end (897K in, 16.7K out) of the OpenHands ranges.

Monthly columns use example volumes. No prompt caching is applied. Coding agents resend a lot of context, so real bills with cache hits are often lower.

Measured cost: a separate 500-task OpenHands study averaged 4.17M tokens and $1.857 per task (arXiv 2604.22750, figures as quoted by Tokenade). At that rate the example volumes come to:

SWE-agent's input range (194K–8.1M tokens) is far wider, so its cost per task varies even more. See also coding agent cost per task and how the harness changes cost. If you pay for a plan rather than the API, compare in the subscription vs API calculator.

One stuck agent can blow past any monthly estimate, so set hard limits as described in how to stop an AI agent loop from running up a huge API bill.

2. Customer support bot

Three separate published workloads:

• Spheron's benchmark: 600 input + 350 output tokens per turn, 3 turns per conversation.
• A production deployment: 847 tokens per conversation, 312 of them output (so 535 input), averaging 12,400 conversations a day. Calls per conversation weren't given, so it's priced as one total.
• Two open-source triage flows on 100 tickets: 4,500 and 6,100 median tokens per ticket, with 3.1 and 4.7 LLM calls per ticket.

Light/Medium/Heavy columns are example volumes. The 12,400/day column uses the deployment's own reported volume.

The triage source reports total tokens per ticket without an input/output split, so this table shows the range from "all input" to "all output" pricing:

3. Research agent

Anthropic's multi-agent research write-up gives the shape of the workload but not absolute tokens per task:

• Simple fact-finding: 1 agent with 3–10 tool calls. Direct comparisons: 2–4 subagents with 10–15 calls each. Complex research: more than 10 subagents.
• Single agents use about 4× the tokens of chat. Multi-agent systems use about 15×.

Gartner, as cited by Cockroach Labs, puts agentic workflows at 10–20 model calls per task and 5–30× the tokens of a chatbot.

To turn the multipliers into dollars, this table uses an example assumption: the chat baseline is the Spheron support conversation above (3 turns × 600 in + 350 out), with the 4× and 15× applied to its cost. That assumes the same input/output mix, which real research agents may not have.

Monthly columns show single-agent (4×) to multi-agent (15×) at example volumes. Swap in your own chat baseline in the calculator.

4. Voice agent

Voice agents are usually billed per minute, not per token, so the token formula doesn't apply. Here is Retell's worked example as published:

• 50,000 calls a month, 3 AI minutes per call, $0.13 per minute with GPT 5.1 ($0.055 voice infrastructure + $0.04 LLM + $0.015 voice + $0.015 telephony + $0.005 knowledge base).
• AI minutes: 50,000 × 3 × $0.13 = $19,500. Telephony after transfer (20% of calls, 6 more minutes at $0.015): $900. Phone numbers: $20. Extra concurrency: $160.
• Total: $20,580 a month, about $0.41 per call. Retell says switching the LLM to GPT 5.5 raises the rate to $0.25 per minute and the total to $38,580.

50,000 calls a month is about 1,667 a day (our arithmetic, not Retell's figure). Prices are Retell's pay-as-you-go rates as of October 2026.

What moves the monthly bill

• Agent type matters more than model. A coding task can use hundreds of thousands of input tokens, while a support conversation uses about a thousand.
• Caching. Repeated context (system prompts, tool schemas, repo files) can be billed at the cached-input price. Model it in the prompt caching calculator.
• Subscriptions. For coding agents, a flat plan can beat API pricing. Check the subscription vs API calculator.
• Multi-agent fan-out. Going from 4× to 15× chat tokens is where research agent costs jump.

FAQ

Run your own numbers

Open the TokenTab calculator Subscription vs API Prompt caching calculator

Related: How to estimate AI agent costs · Cheapest LLM for AI agents · MCP tool token cost.