The short answer
For a typical small business agent, budget $50 to $500 a month to run it. That covers model API usage, hosting, retrieval and logging. Most of that is model usage, and model usage comes down to three numbers: tokens per conversation, conversations per month and the price of the model tier you pick.
Low-volume internal agents often sit at the bottom of that range. A customer-facing agent handling tens of thousands of conversations on a top-tier model can cost thousands a month in model usage alone, so forecast volume before you choose a model.
How the model bill is calculated
Model providers charge per million tokens, with separate prices for input (what you send) and output (what the model writes). A token is roughly three quarters of an English word.
What surprises most teams is that an agent does not make one call per conversation. It makes one call per step. A support agent that looks up an order, checks a policy and then replies makes three or four calls, and each call resends the system prompt, the tool definitions, the conversation so far and any retrieved documents. That is why input tokens usually dominate the bill even though output tokens cost more each.
On current Claude models, reasoning (thinking) tokens are billed as output. Higher effort settings produce more of them.
A worked example with Claude list prices
Anthropic’s published API list prices as of October 2026, per million tokens, are: Claude Opus 5.5 at $4 input and $20 output, Claude Sonnet 5.5 at $2 input and $10 output, and Claude Haiku 5.5 at $0.10 input and $0.50 output. These are list prices; confirm them on Anthropic’s pricing page before you budget, because they change. Haiku’s listed rate applies to shorter prompts, and very long prompts are billed at a higher rate.
Assume a support agent where one conversation adds up to 20,000 input tokens and 1,500 output tokens across all its steps:
| Opus 5.5 | Sonnet 5.5 | Haiku 5.5 | |
|---|---|---|---|
| Input cost per conversation (20,000 tokens) | $0.080 | $0.040 | $0.002 |
| Output cost per conversation (1,500 tokens) | $0.030 | $0.015 | $0.00075 |
| Total per conversation | $0.110 | $0.055 | about $0.003 |
| 2,000 conversations a month | $220 | $110 | about $5.50 |
| 20,000 conversations a month | $2,200 | $1,100 | about $55 |
At 2,000 conversations a month, a Sonnet-based agent spends about $110 on the model. Add hosting, a vector store and logging and it lands comfortably inside the $50 to $500 range. Move the same agent to Opus for every step and the model line roughly doubles. Grow to 20,000 conversations and you are in a different budget conversation altogether.
Where the rest of the money goes
| Cost line | What it pays for | How it scales |
|---|---|---|
| Model usage | Input, output and reasoning tokens | Linearly with conversations and context size |
| Retrieval | Embeddings for your documents, a vector database or pgvector, extra input tokens per call | With document volume and passages pulled per call |
| Hosting | The server or functions that run the agent loop and tools | Mostly flat at small volume |
| Monitoring | Logs, traces, evaluation runs, alerting | With call volume and how much you retain |
| Telephony or messaging | Phone minutes, SMS or WhatsApp fees for voice and messaging agents | Per minute or per message, billed by the carrier or platform |
Voice agents add per-minute telephony and speech costs on top of the model, which is why they cost more to run than a chat agent at the same volume. See what a voice AI agent costs for that build.
How to keep running costs down
- Route by difficulty. Use Haiku 5.5 for classification, routing and simple extraction, Sonnet 5.5 for most reasoning, and Opus 5.5 only on the steps that fail without it. Claude Fable 5.1 is the top tier and costs more again, so reserve it for problems that genuinely need it.
- Turn on prompt caching. If 12,000 of those 20,000 input tokens are the same system prompt and tool list on every call, caching bills them at a fraction of the normal input price. Keep the stable content first and the changing content last so the cache actually hits.
- Trim what you resend. Summarize long histories, retrieve fewer and tighter passages, and remove tools the agent never uses on a given route.
- Tune effort and output length. Lower effort settings and a clear instruction on answer length cut output and reasoning tokens with little quality loss on routine work.
- Batch what can wait. Overnight jobs such as report summaries can use the provider’s batch processing, which is billed at a discount.
- Cap the loop. Limit steps per conversation and alert on daily spend so a bug cannot run up a large invoice.
If your data or volume points toward self-hosting a model instead, read open-source LLMs vs API models first. Self-hosting trades per-token fees for GPU and engineering costs, and it rarely wins at small business volume.
How Lumeor estimates running costs
Before we build an agent, we estimate tokens per conversation from the actual prompt, tools and retrieval design, multiply by your expected volume, and show the monthly figure for each model tier you might use. After launch, we log tokens per call so the estimate becomes a measured number. The build itself is priced separately; see what a custom AI agent costs. If you want help sizing your own agent, start here.