How Do You Estimate the Monthly Cost of an LLM Step in a Workflow?

How much will that AI step in your workflow actually cost each month?

Multiply three numbers: how many times the step runs per month, how many tokens each run uses, and the model's price per token. Do it separately for input and output tokens, because output usually costs more. Add a buffer for retries. Most teams can estimate an LLM step within a reasonable range in under an hour.

AI steps have become a normal part of automations. A Zapier or Make workflow summarizes a form submission, an n8n flow classifies inbound leads, or a Clay table writes a first line for outbound. Each of those steps calls a language model, and each call costs money.

The cost is usually small per run, which is exactly why it gets ignored. Then volume grows, someone switches to a bigger model, or a loop starts retrying, and the monthly bill surprises everyone. I build and run AI automations for B2B teams, and I estimate every LLM step before it ships. Here is the method.

What determines the cost of an LLM step?

Four things: the number of runs, the input tokens per run, the output tokens per run, and the model's prices for input and output. Tokens are chunks of text that models read and write. Input includes your instructions, any context, and the data you pass in. Output is what the model writes back.

Most model providers price input and output separately, quoted per million tokens. Output almost always costs more than input. That means a step that writes long answers can cost far more than one that reads a long document and returns a short label.

A rough rule many people use is that a token is around three quarters of an English word, though it varies by language and content. For estimating, that is close enough. For precision, most providers offer a way to count tokens on real examples, so check your provider's documentation.

Platform costs sit on top. If the LLM call runs inside Zapier, Make, n8n Cloud, or Clay, those platforms may charge for the task or credit as well. Check each platform's own pricing page, because those costs can matter as much as the model.

How do you count the tokens in one run?

Take three to five real examples of the input the step will see, build the full prompt for each, and count the tokens with your provider's tokenizer or token counting tool. Then run the step and record the output length. Use the average, and note the largest example, because outliers drive surprise bills.

Remember everything that goes into the prompt. The instructions you wrote, any examples you included, the system message, and the data from earlier steps all count as input. Long instructions repeated on every run add up quickly.

Watch for hidden context. If the step passes a whole CRM record or a full email thread when it only needs two fields, you are paying to send data the model does not use. Trimming inputs is often the cheapest optimization available.

For output, set a sensible maximum length and ask for concise formats. A classification step should return a label, not a paragraph explaining the label.

What does a worked example look like with current prices?

Take a hypothetical step that sends 2,000 input tokens and gets 500 output tokens back, running 3,000 times a month. At the Claude Sonnet 5.5 prices on Anthropic's pricing page today, that is about $27 a month. On Claude Haiku 5.5, the same workload is about $1.35. Model choice dominates.

Here is the math. Anthropic's pricing page lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens. Three thousand runs at 2,000 input tokens is 6 million input tokens, which is $12. Three thousand runs at 500 output tokens is 1.5 million output tokens, which is $15. Together, $27.

The same page lists Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Six million input tokens is $0.60, and 1.5 million output tokens is $0.75. Together, $1.35.

These numbers are an illustration, not a quote. Prices change, and they differ across providers and plans. Always check the provider's current pricing page before you commit to a budget.

How do model choice and step design change the bill?

Model choice usually has the largest effect, followed by prompt length and output length. Many routine automation tasks, like classifying a lead, extracting fields, or tagging a ticket, work well on smaller, cheaper models. Save larger models for steps where quality clearly improves outcomes, and test that claim on real examples.

I wrote about this trade-off when I looked at which GTM jobs should move to a small model. The short version is that many steps were built on a large model by default, not by decision.

Step design matters too. Splitting one big prompt into a cheap filtering step followed by an expensive step that only runs on the records that pass can cut costs sharply. If only a fifth of inbound leads need a detailed summary, only a fifth should pay for one.

Caching and batching can help where they fit. Anthropic's pricing page, for example, says its Batch API processes large volumes of requests asynchronously with a 50% discount on both input and output tokens. Batching only suits steps that do not need an instant answer, like overnight enrichment.

Why do real costs often exceed the estimate?

Retries, loops, and volume growth. A failing step that retries several times multiplies cost. A loop that processes the same records twice doubles it. And the volume you estimated for this month is rarely the volume you will have in six months. Build a buffer into the estimate and monitor actual spend from day one.

Retries deserve special attention. When an automation platform or your own code retries a failed LLM call, each retry is billed. If a malformed input keeps failing, you can pay many times for one record. My notes on handling rate limits and retries in a marketing automation cover how to cap that.

Prompt drift is another quiet cause. Someone adds a few examples to improve quality, and the input length triples. Nobody re-runs the estimate. Treat prompt changes like code changes, and recheck the cost when the prompt grows.

Finally, test runs count. Development and debugging can consume a surprising amount of tokens, especially when testing on large datasets. Budget for it.

How should you monitor LLM costs after launch?

Set a spending limit or alert in your provider's console, log token usage per run where your platform allows it, and review actual cost against your estimate after the first week and the first month. If actual cost is far above the estimate, find out which assumption broke before you scale the workflow further.

Most model providers offer usage dashboards and some form of spending limits or alerts. Set them before launch, not after the first surprising invoice. Check your provider's documentation for what is available on your plan.

Track cost per useful outcome, not just total cost. If an AI step costs a few dollars a month and saves hours of manual review, the cost is trivial. If it costs a little but nobody uses the output, any cost is too much.

Put the estimate and the actuals in your automation inventory. A simple column for expected monthly cost and another for last month's actual cost makes drift visible during regular reviews.

When is an LLM step not worth the cost?

When a simple rule would do the job just as well. If a field can be checked with a filter, a lookup, or a regular expression, using a language model adds cost, latency, and unpredictability for no gain. Use AI where judgment or language understanding is actually needed, and plain logic everywhere else.

A good test is to write the rule you would follow if you did the task by hand. If that rule fits in one sentence with no exceptions, it probably belongs in normal automation logic. If it needs judgment, context, or reading between the lines, an LLM step may earn its keep.

I also look at the cost of being wrong. A cheap step that occasionally mislabels a lead might cost more in lost deals than an expensive step that rarely does. The cheapest model is not always the cheapest choice.

What should you do next?

Pick one AI step in your workflows. Count runs per month, measure tokens on five real examples, and multiply by the current prices from your provider's pricing page. Add a buffer for retries and growth. Then set a spending alert and compare actual cost with your estimate after the first month.

That hour of math turns an invisible cost into a managed one, and it often reveals a cheaper model or a shorter prompt that works just as well.

If you want help estimating, trimming, or monitoring the AI steps in your automations, reach out at pravinkumar.co. Let's chat.