Giacomo Mantovani

Article · November 10, 2026 · By Giacomo Mantovani

What an AI system costs to run each month: the line items to budget for

The build has a quote; the monthly bill does not. The five line items behind the running cost of an AI system in production, and what pushes each up.

  • AI costs
  • AI in production
  • LLM
  • budgeting
The build has a quote. The monthly bill doesn't. Five line items, Tokens −80% after the rewrite, Same features.

Most AI projects come with a clear build estimate. The surprise comes after launch, when the monthly bill grows with usage and sometimes faster than usage. On a consumer finance and investing platform where I owned the AWS backend, the inherited AI pipelines made more model calls than they needed, sent oversized context and had no ceiling on spend. Rewriting them cut AI costs by about 80%. The features and the model stayed the same.

What goes into the monthly cost of running an AI system?

There are five line items, and only the first one is set by your model provider. The other four depend on how the system is built and on how much human work is left around it.

Monthly running cost line items
Line itemWhat it pays forWhat pushes it upHow to control it
Model tokensInput and output text of every callDuplicate calls, whole documents in prompts, repeated generationsCost logged per call and per feature, caching, spend caps
InfrastructureServers, job queues, databases, storageVolume spikes, slow processing, environments left runningQueues with workers added only when load needs them
Monitoring and logsTracking calls, errors, cost and qualityLogging full text, long retentionTargeted logs, no sensitive data in logs
Human reviewPeople checking uncertain casesFew automated checks, a clumsy review screenDeterministic checks, review of exceptions only
MaintenancePrompts, new model versions, quality testsNo test set, model swaps judged by eyeVersioned prompts and a fixed evaluation set

How much of the AI bill is model tokens?

Per-token prices differ a lot between models. On Anthropic's price list, Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens, while Claude Opus 5.5 costs $4 and $20, forty times more (Anthropic, October 2026). For both, output costs five times as much as input, so long answers cost more than long prompts.

The token price still tells you little about your monthly spend. What matters is how many calls the system makes per result and how much text each call carries. On the finance platform, every stock report is a 13-slide deck with KPI extraction, source-domain checks, output verification and Italian translation. The 80% saving kept every one of those steps. What went was redundant calls and context the model never used. The full story is in the LLM cost reduction case study.

Why do AI costs grow faster than usage?

A system that started as a prototype reaches production with prototype habits: the same question sent to the model twice, the whole conversation history attached to every message, content regenerated on every request instead of stored. With ten users nobody notices. With a thousand users every bit of waste is multiplied by a thousand, and an automated job with no limit can run all night because of a bug.

It can go the other way too. In Estyl's fashion assistant, which searches a catalogue of more than 2 million products, latency and token spend per request went down while usage grew. Its product retrieval combines hybrid search, structured filters and re-ranking. A pipeline that sends the model only the data it needs costs less per request, so the bill grows more slowly than the user base.

How much does the infrastructure around the model cost?

Mostly it depends on volume and on how that volume arrives. An assistant answering a few hundred questions a day fits on a small service. A high-volume document pipeline needs queues and parallel workers. The one I designed for financial statement extraction processes more than 2,500 documents a day with 16 workers behind Kafka on the client's OpenShift. Each document is an independent message on the queue, so you add workers to handle more volume and the cost follows the load.

Where the system runs matters as much as volume. For the bank clients, the model ran on Azure OpenAI inside the client's EU cloud and the pipeline on their own platform. Part of the infrastructure was already paid for, but security reviews, audit trails and compliance became work that needed a budget line. Location can change the token price as well: on Anthropic's API, pinning inference to the US adds a 1.1x multiplier on every token category for recent models (Anthropic, October 2026).

Should human review be in the monthly budget?

Yes, and it is often the biggest line. An extraction system with no checks leaves the work where it was, because someone has to recheck everything. In the financial statement pipeline every document goes through about 30 automated checks, such as totals that must reconcile. Documents that pass all of them go through with nobody touching them. The rest reach an analyst with the failing check already named. Manual work fell by 84%.

To estimate this line you need the share of cases that pass automatically. Measure it on a sample of real cases before launch, then every month in production.

How do you keep monthly AI spend under control?

I work in the same order every time. First, measure: log the cost of every call and every feature. On the finance platform each generated report stores its own cost. Second, remove redundant calls, unneeded context and repeated generations. Third, cache answers that do not need regenerating. Fourth, set limits in code: AI images for news articles have daily and per-category spend caps, and an image is generated only for articles above a score threshold. Model choice comes last.

Work that does not need a real-time answer, such as overnight reports or reprocessing an archive, can run at a discount. On Anthropic's Batch API input and output tokens cost 50% less, and repeated text read from the prompt cache costs 10% of the normal input price, or 5% on Opus 5.5 and Sonnet 5.5 (Anthropic, October 2026).

What mistakes push the monthly cost of an AI system up?

The most common is choosing a model by its per-token price when spend actually depends on the number of calls and the amount of text sent. Close behind is launching without spend caps, so a buggy automated job can burn a month's budget overnight. Then there is tracking cost only as the invoice total, which leaves you with no idea what to optimise. Many teams leave human review out of the budget and find out after launch that staff recheck everything. The last is switching models to save money without a test set, and learning about the quality drop from customer complaints.

Next step

Want to know what your AI system really costs to run?

I trace where your tokens go and tell you which line items can come down without losing features.

Let's talk

If your bill is growing faster than usage, the LLM cost optimisation service page explains how I run the audit.

Frequently asked questions

Is switching to a cheaper model enough to cut costs?

Not necessarily. On the finance platform the saving of about 80% came from rewriting the pipelines (fewer calls, smaller context, caching, spend caps), not from changing the model. Fix the architecture first, then decide whether a model change is still worth it.

Can I estimate monthly AI running costs before launch?

Roughly, yes. Run a few hundred real cases, measure tokens and calls per result, multiply by expected volume and add a margin for peaks. In the first month of production, compare the estimate with the logged cost.

Does cutting AI costs lower answer quality?

It shouldn't, as long as the checks stay. On the finance platform, reports still verify source domains and check the model's output. A fixed evaluation set, run before and after every change, confirms that quality has not dropped.

Is self-hosting a model cheaper than paying per API call?

It depends on volume and data constraints. A self-hosted model has a fixed hardware and operations cost that pays off only at high, steady volume. If the constraint is keeping data in a region, a cloud service in that region is often enough: for the bank clients the model ran on Azure OpenAI in the client's EU cloud.

Author

Giacomo Mantovani

AI Engineer & Consultant for Companies and Agencies

LinkedInGitHub

Next step

Tell me the workflow. We scope it on a free call, then a working version is live within 10 days.

Book a scoping call

Related articles