How much does an AI system cost to run each month in 2026?

Provider prices are USD list prices checked on each pricing page on 7 November 2026. Agency prices are in GBP as each agency publishes them.

The four bills below use the prices above and stated volumes. Token counts are assumptions, so measure your own in a pilot before you set a budget.

Support agent that answers from your help centre

Assume 2,000 conversations a month, each with 3,000 input tokens (the question, the instructions and the help articles retrieved for it) and 500 output tokens.

• On Claude Sonnet 5.5, 3,000 × $2 ÷ 1,000,000 = $0.006, plus 500 × $10 ÷ 1,000,000 = $0.005. That is $0.011 a conversation, or $22 a month.
• On gpt-6.1-terra, $0.003 + $0.0025 = $0.0055 a conversation, or $11 a month.
• On Claude Haiku 5.5, $0.0003 + $0.00025 = $0.00055 a conversation, or $1.10 a month.

Search over 1,000 help-centre chunks of 1,536 dimensions stores 1.5 million dimensions and, by Cloudflare's formula, queries (1,000 + 2,000) × 1,536 = 4.6 million a month, inside the Workers Paid allowance. On Sonnet 5.5 the bill is $22 + $5 hosting + $29 monitoring = $56 a month.

Inbox triage agent

Assume 3,000 emails a month, about 150 a working day. The agent sorts every email and drafts a reply for 900 of them (30%) for a person to approve.

• Sorting on Claude Haiku 5.5, at 1,500 input and 50 output tokens an email, costs $0.00015 + $0.000025 = $0.000175 an email, or $0.53 for 3,000.
• Drafting on Claude Sonnet 5.5, at 4,000 input and 400 output tokens a draft, costs $0.008 + $0.004 = $0.012 a draft, or $10.80 for 900.
• Model usage totals $11.33, and the bill is $11.33 + $5 + $29 = $45.33 a month. Sorting on Sonnet 5.5 too would cost $10.50 instead of $0.53.

Document processor for supplier invoices

Assume 1,000 single-page scanned invoices a month. Anthropic's vision docs charge one token per 28 × 28 pixel block, so an A4 scan at 150 dpi (1,240 × 1,754 pixels) is 45 × 63 = 2,835 tokens. With instructions and a field list that is about 3,500 input tokens, plus 500 output tokens of extracted fields.

• On Claude Sonnet 5.5, $0.007 + $0.005 = $0.012 a page, or $12 a month. Batch processing halves that to $6.
• On Claude Haiku 5.5, $0.00035 + $0.00025 = $0.0006 a page, or $0.60 a month.

Storing 1,000 PDFs at an assumed 200 KB each takes 0.2 GB, inside the 10 GB-month that Cloudflare R2 gives free. On Sonnet 5.5 in batch mode the bill is $6 + $5 + $29 = $40 a month. The AI document processing guide compares this with Dext and AutoEntry.

Voice agent on a UK phone line

Assume 1,000 inbound calls a month at 3 minutes each, which is 3,000 minutes.

• Retell's voice agent costs 3,000 × $0.07 to $0.31 = $210 to $930.
• UK telephony on a number bought through Retell costs 3,000 × $0.10 = $300.
• The number costs $2 a month.

That is $512 to $1,232 a month, or $517 to $1,237 with $5 of hosting for the booking tools. Retell's range depends on the model and voice, and its voice infrastructure alone costs $0.055 a minute. The voice agent or phone menu guide compares this with a simple menu.

The four bills side by side

All figures are USD, before VAT and before any care plan. The voice bill is 9 to 22 times the support agent's, because Retell bills every minute of every call.

UK agency estimates run higher than this token maths. Augustova's pricing page says a first agent or automation typically costs £50 to £250 a month to run, and Softomate's chatbot guide puts AI usage at £20 to £80 a month for 500 conversations. Ask any supplier for the token counts behind its estimate.

Four habits lower the bill. Send routine steps to a small model, cache text that repeats, batch work that can wait and trim what goes into each prompt.

Use a smaller model for routine steps

Claude Haiku 5.5 costs $0.10 per million input tokens against $2 for Sonnet 5.5, a 20-fold gap, and OpenAI's gpt-6-luna costs the same as Haiku. In the inbox example, sorting on Haiku costs $11.33 a month in total against $21.30 with Sonnet for both steps, a 47% saving. Test the small model on 50 to 100 real inputs first.

Cache the parts that repeat

Cached input costs 90% to 95% less than fresh input at Anthropic and OpenAI, for example $0.10 against $2 per million on Sonnet 5.5 and $0.05 against $1 on gpt-6.1-terra. Take an assistant that sends a 20,000-token price list and policy document with each of 2,000 questions a month.

• Without caching, that document costs 2,000 × 20,000 × $2 ÷ 1,000,000 = $80 a month.
• With Anthropic's 5-minute cache, assume one question in five arrives after the cache has expired. 400 cache writes at $2.50 per million cost $20, and 1,600 cache reads at $0.10 per million cost $3.20.
• The total is $23.20, a saving of 71%.

The cache pays only when questions arrive close together.

Batch work that can wait

Anthropic, OpenAI and Google all charge half price for batch requests, so gpt-6.1-terra falls to $0.50 input and $2.50 output per million tokens. Invoice reading and overnight reports suit batches. Live chat and phone calls don't.

Trim each prompt and watch the price calendar

Input tokens outnumber output tokens in every text example above, so retrieving three passages instead of ten cuts the bill directly. On voice, Retell's $0.055 infrastructure and $0.10 UK telephony apply to every minute whatever the model, so call length drives the bill. Google's gemini-4-argon doubles on 1 December 2026, which would take the support agent from $22 to $44 a month.