# What an AI Agent Costs to Run After the Build > The monthly bill after an AI agent is built has four lines: the model, the messaging channel, hosting and care. Where each comes from, with vendor price pages. - Web page: https://sporesec.com/en/blog/ai-agent-running-costs - Language: English - Last updated: 2026-10-30 - Also in עברית: https://sporesec.com/he/blog/ai-agent-running-costs.md ## Answer The monthly bill after an AI agent is built has four lines: the model, the messaging channel, hosting and care. Where each comes from, with vendor price pages. - By Mahdy Lorens Gribkov - Published: 2026-10-30 - Pricing, 5 min read - Keywords: AI agent running cost, how much does an AI agent cost, LLM API pricing small business, AI agent monthly cost After an AI agent is built, the monthly bill has a few separate lines: the language model, the messaging channel, hosting, and the care that keeps the agent working. For a small business with modest volume, the model line can be small. The fixed lines and the people lines are usually what you actually budget for. This post shows where each line comes from, with the vendors' own price pages, so you can ask a vendor to show their numbers. ### What does the language model cost per conversation? The model is billed by tokens, small pieces of text, with a separate price for what goes in and what comes out. Anthropic's [pricing page](https://platform.claude.com/docs/en/about-claude/pricing) (read 2026-10-28) lists, per million tokens: - Claude Haiku 4.5: $1 input, $5 output. - Claude Sonnet 5.5: $2 input, $10 output. - Claude Opus 5.5: $4 input, $20 output. - Claude Haiku 5.5, a smaller and cheaper model the same table also lists: $0.10 input, $0.50 output, for prompts up to 100,000 tokens. Here is an illustration, and it is our arithmetic on those rates, not a figure from Anthropic. Take a support conversation of 3,700 tokens, which we assume splits into 3,000 going in and 700 coming out. On Haiku 4.5 that is about $0.0065, or about $65 for 10,000 conversations. On Sonnet 5.5 it is about $0.013, or $130 per 10,000. On Opus 5.5 it is about $0.026, or $260 per 10,000. On Haiku 5.5 it is about $0.00065, or $6.50 per 10,000. Two cautions apply. The page says that Claude 4.7 and later models use a tokenizer that produces roughly 30% more tokens for the same text, so for the newer models treat the figures above as a floor. And your real conversations may be longer than 3,700 tokens, because the agent re-reads its instructions and the history each turn. Three levers from the same page change the number: - **Prompt caching.** Reading cached input costs one tenth of the normal input price on most models, and one twentieth on Opus 5.5. That matters when the agent re-sends the same long instructions every turn. - **Batch processing.** Work that does not need an instant answer, such as overnight summaries, gets a 50% discount. - **Tools.** Using tools adds tokens (the tool definitions and results count as input), and web search is billed separately at $10 per 1,000 searches. ### What does the messaging channel cost? If the agent talks to customers on WhatsApp, Meta charges per message delivered, on top of the model. Two details changed recently. Meta's page on [non-template messages](https://developers.facebook.com/documentation/business-messaging/whatsapp/pricing/non-template-messages) (read 2026-10-28) says that since 1 October 2026 Meta charges for service messages, the free-form replies to a customer who wrote to you first, at the same rate per market as utility and authentication messages. They had not been charged since November 2024. Meta's [pricing page](https://developers.facebook.com/documentation/business-messaging/whatsapp/pricing) says that messages inside the free entry point window, opened when you answer a customer who came from a Click to WhatsApp ad, are not charged, and that the window may stay open for up to 7 days. Rates differ by country and are published as a downloadable rate card, where Israel has its own line. We do not quote an Israel figure here, because the page itself does not show one. Read the current rate card for Israel before you budget, and multiply by your expected message count. A conversation is several messages, so count messages, not conversations. For how this fits a full bot price, see [what a WhatsApp bot costs in Israel](https://sporesec.com/en/blog/whatsapp-bot-pricing-israel). ### What does hosting cost? Small agents run on small infrastructure. As one example, Cloudflare's [Workers pricing page](https://developers.cloudflare.com/workers/platform/pricing/) (updated 2 October 2026, read 2026-10-28) lists a Paid plan with a minimum charge of $5 per month, including 10 million requests and 30 million CPU milliseconds, with overage at $0.30 per additional million requests. A free plan exists with a limit of 100,000 requests a day. The point is not that this particular platform is the answer. It is that hosting for a small agent is typically a minor, predictable line, and a vendor should be able to name it. ### What is not on any price page? The lines that are not on a vendor's price page are the ones that decide whether the agent stays useful: - **Monitoring and tuning.** Someone has to read the conversations that went wrong, fix the instructions and update the knowledge as your business changes. - **Human handoff.** When the agent hands a conversation to a person, that person's time is a cost, and so is the lead lost if no one answers. - **Change requests.** New services, new prices, a new flow. Ask your vendor whether these are included in a monthly fee, quoted as a fixed price per change, or left to you. With us, care is optional and is quoted together with the build. ### What to do this week 1. Estimate your volume: conversations per month and messages per conversation, from your current inbox or phone log. 2. Ask the vendor for a line-by-line monthly estimate: model, messaging channel, hosting, care. Check each line against the price pages linked below. 3. Ask what happens to the bill if volume triples, and who gets alerted if usage spikes. ### Sources - [Pricing, Anthropic Claude API documentation](https://platform.claude.com/docs/en/about-claude/pricing), read 2026-10-28. - [Non-template message charging, Meta WhatsApp Business Platform](https://developers.facebook.com/documentation/business-messaging/whatsapp/pricing/non-template-messages), read 2026-10-28. - [Pricing, Meta WhatsApp Business Platform](https://developers.facebook.com/documentation/business-messaging/whatsapp/pricing), read 2026-10-28. - [Pricing, Cloudflare Workers documentation](https://developers.cloudflare.com/workers/platform/pricing/), updated 2 October 2026, read 2026-10-28. ### Related - [AI agents for small businesses](https://sporesec.com/en/services/ai-agents) - [What a WhatsApp bot costs in Israel](https://sporesec.com/en/blog/whatsapp-bot-pricing-israel) ## Questions and answers ### How much does the language model cost per conversation? On Anthropic's published rates, a conversation of 3,000 input and 700 output tokens costs about 0.65 US cents on Claude Haiku 4.5, about 1.3 cents on Sonnet 5.5 and about 2.6 cents on Opus 5.5. The token split is our assumption, and real conversations are often longer. ### Does WhatsApp charge for an AI agent's replies? Since 1 October 2026 Meta charges for service messages, the free-form replies to a customer who wrote first, at the same per-market rate as utility messages. Messages inside the free entry point window, opened by Click to WhatsApp ads, stay free. ### What costs are not on a vendor's price page? Monitoring and tuning, the human time spent on handoffs, and change requests. Ask whether they are in a monthly fee, quoted as a fixed price per change, or left to you. ## Next step - Book a free intro call: https://sporesec.com/en/book.md - Blog: https://sporesec.com/en/blog.md --- - Contact: contact@sporesec.com | WhatsApp https://wa.me/972532322318 | https://sporesec.com/en/contact.md - Book a free intro call: https://sporesec.com/en/book.md - Machine-readable site data: https://sporesec.com/llms.txt | https://sporesec.com/agents/en/index.json