What an AI Agent Actually Costs to Run Each Month — The Token Arithmetic

For most business agents running at Indian SMB volumes, the monthly model bill comes to somewhere between a few hundred and a few tens of thousands of rupees. Which end you land at is decided mostly by one choice — which model tier you run — and secondarily by how many steps each task takes. It is rarely the line that kills a project.

What does kill projects is not knowing the number in advance, then discovering in month two that a handful of runs cost fifty times the average. This post shows how to work the number out yourself, from the providers' own price lists, before anyone signs a build quote.

We have written separately about what agents can genuinely automate and the difference between automation and an agent. This is the part those posts skip: the meter.

How is an AI agent billed?

By the token, on both sides. Every time the agent calls the model, you pay for everything you send in (input tokens) and everything the model writes back (output tokens), at separate rates. Output is always the expensive side — usually five times the input rate.

Here are current list prices, in US dollars per million tokens, taken from each provider's own pricing page as of this week:

Model	Input	Output	Tier
Claude Haiku 5.5	$0.10	$0.50	Small, fast
Claude Sonnet 5.5	$2	$10	Mid
Claude Opus 5.5	$4	$20	Large
Gemini 3.8 Flash	$0.75	$3.75	Mid (rising in January)
GPT-6 Luna	$0.10	$0.50	Small, fast
GPT-6 Astra	$10	$50	Flagship

Sources: Anthropic's pricing documentation, Google's Gemini API pricing, OpenAI's API pricing. Claude Haiku 5.5's rate applies to prompts up to 100,000 tokens; above that it rises fivefold.

Two warnings before you use that table. First, a token is not the same unit across providers. Anthropic's own documentation notes that its newer tokenizer "produces approximately 30% more tokens for the same text" than its older one, so per-token prices are only comparable within one vendor. Second, prices move, and not only downwards. Google's page lists Gemini 3.8 Flash doubling to $1.50 input and $7.50 output from 1 January 2027, and its flagship doubling from 1 December. Any cost model you build needs a date on it.

Why does an agent cost more than a chatbot?

Because it calls the model repeatedly, and it resends the whole conversation every time.

A chatbot answers a question: one call, done. An agent works a task — read the enquiry, look up the customer, check stock, draft a reply, log it — and each step is a fresh call that carries the instructions, the tool definitions and everything that has happened so far. The context grows with each step, so the cost of a run grows faster than the number of steps.

Anthropic, writing about its own research agent, reported that "agents typically use about 4× more tokens than chat interactions," and multi-agent systems about 15×. That is one company's system, not a law, but it is the right order of magnitude to carry in your head.

There is also overhead you never see. Simply giving a Claude model access to tools adds a hidden system prompt — 286 tokens on the current models, per the pricing page — before your own instructions and tool descriptions, which in a real agent typically run to a few thousand tokens. You pay for those on every single call.

A worked example: an enquiry-handling agent

Take a common first project. An agent reads an inbound enquiry, looks the sender up in the CRM, checks two facts, drafts a reply for a human to approve and logs the outcome. Assume:

• 4,000 tokens of fixed instructions and tool definitions
• 500 tokens for the enquiry itself
• 5 model calls per run, each adding about 600 tokens of tool results and reasoning to the context
• 200 output tokens per call

That gives roughly 28,500 input tokens and 1,000 output tokens per run. At 100 enquiries a day, that is about 3,000 runs a month. Converting at ₹90 to the dollar — a round figure for illustration, so use the day's rate for your own version:

Model	Per run	Per month (3,000 runs)
Claude Haiku 5.5	₹0.30	about ₹900
Gemini 3.8 Flash	₹2.26	about ₹6,800
Claude Sonnet 5.5	₹6.03	about ₹18,100
Claude Opus 5.5	₹12.06	about ₹36,200
GPT-6 Astra	₹30.15	about ₹90,500

The spread is a factor of a hundred between the cheapest and the most expensive row, for the same task. That is the single most important thing in this post: the model tier you choose matters far more than anything else you tune.

It also says something about where the money should go. For a triage-and-draft job like this, a small model is frequently good enough, and the difference between ₹900 and ₹90,000 a month buys a lot of evaluation work to prove it. Test the cheap tier first, against real enquiries, and move up only where it demonstrably fails.

These are list prices before tax. If you are a GST-registered business paying a foreign provider directly, the CBIC's guidance on online services is that the tax is paid by you under reverse charge, which is normally claimable back as input credit. Your CA will want to know it is happening.

What makes the bill spike?

The table above is an average. Agents do not produce averages; they produce distributions with a long tail. Three things drive the tail.

Loops. An agent that cannot find what it needs keeps trying. Our example run at 5 steps costs ₹6 on Sonnet 5.5. The same run at 25 steps — because a CRM lookup kept failing and the agent kept rephrasing — costs about ₹57, not ₹30, because each extra step carries the growing context with it. Ten such runs a day is invisible in the first week and very visible on the invoice.

Long inputs. Attach a 40-page PDF, a long email thread or a full product catalogue to every call and the input side dominates. On models with a long-context surcharge, such as Haiku 5.5 above 100,000 tokens, you also cross into a higher price bracket for the whole request.

Paid tools. Web search is billed per search on top of tokens — $10 per 1,000 searches on both Anthropic's and OpenAI's APIs, about ₹0.90 each at our illustrative rate, and the results then count as input tokens. An agent that searches five times per run adds about ₹4.50 to our ₹6 Sonnet example before it has read a single result.

The practical defences are dull and effective: a hard cap on steps per run, a hard cap on spend per run, an alert when the daily total crosses a threshold, and logging the cost of every run so you can see the 95th percentile, not just the mean.

How much do caching and batching actually save?

More than most people expect, on the right workload.

Prompt caching lets you pay once for the fixed part of every request — the instructions and tool definitions — and then read it back cheaply. On Anthropic's API a cache write costs 1.25× the input price for a five-minute cache, and a cache read costs 0.1× normally, or 0.05× on Sonnet 5.5 and Opus 5.5. Google and OpenAI both list similar cached-input rates.

Applied to our example, caching only the 4,000-token fixed prefix brings Sonnet 5.5 from about ₹6.03 per run to about ₹3.47, and the monthly figure from roughly ₹18,100 to ₹10,400. That is a configuration change, not a rebuild. The catch is that caches expire — five minutes or an hour on Anthropic's API — so the saving depends on traffic arriving often enough to keep the cache warm.

Batch processing halves the price on all three providers in exchange for not getting the answer immediately. Anthropic, Google and OpenAI each list it at 50% of standard rates. It is useless for anything a customer is waiting on, and ideal for overnight work: reclassifying last month's tickets, enriching a lead list, summarising the day's calls. Anthropic's documentation notes the two discounts can be combined.

Lever	Typical saving	Works for	Does not work for
Smaller model tier	Up to 100×	Triage, extraction, routing, drafting	Hard reasoning, ambiguous judgement
Prompt caching	About half the input side, in our example	High-volume, same instructions every call	Sporadic traffic that lets the cache expire
Batch API	50%	Overnight and bulk jobs	Anything a person is waiting for
Step and spend caps	Removes the tail	Every production agent	—

What this does not include

The model bill is one line. The others are real, and some are larger:

• Hosting and the queue that runs the agent, plus any database, logging and monitoring. Usually modest for SMB volumes, but not zero.
• The services it calls — WhatsApp conversation charges, SMS, enrichment APIs, a paid CRM tier with API access. These frequently cost more than the model. We broke down WhatsApp's charges in India separately.
• Human review time, which is the cost most underestimated in agent projects. If a person approves every draft, the agent is saving typing, not headcount, and the business case should say so.
• Maintenance, because models get retired and replaced. Anthropic's pricing page lists several older models as retired already. Each migration means re-testing the prompts.

When not to build an agent at all

If your arithmetic shows that a deterministic workflow — no model, or one model call to classify and then fixed rules — would do the job, build that. It costs effectively nothing per run, never loops, and fails visibly. Our own rule of thumb is that most things sold as agents should be workflows with one smart step in them, and the running cost difference is one of the clearest reasons why. Our automation pricing covers what those cost to build.

And do not let a vendor quote you a running cost without showing the inputs: tokens per run, steps per run, model, volume, and what happens on the worst run. A single monthly figure with no working behind it is a guess.

How to estimate yours in an afternoon

1. Write down one real task, step by step, as the agent would do it.
2. Count the tokens in your instructions and tool definitions using the provider's token counter, not a word count.
3. Estimate steps per run, then double it for the tail.
4. Multiply by your real monthly volume, at the cheapest tier that might plausibly work.
5. Run 50 real cases through that tier and check the actual usage the API reports against your estimate.

Step five takes a day and replaces every number in this post with your own. It is the most useful day of the whole project.

If you would like a second opinion on the arithmetic — or on whether the job needs an agent at all — tell us what the task is. We will show you the working, and if a ₹900-a-month model or a plain workflow does the job, that is what we will recommend.