What an AI Agent Actually Costs to Build and Run in Malaysia

The question arrives in almost the same form every time: how much for an AI agent? And the honest first answer is that the question contains three different costs which behave nothing like each other, and most quotations only cover one of them.

There is what it costs to build. There is what it costs to run, which is the token bill and is almost always the smallest of the three. And there is what it costs to keep running, which is the one nobody puts in a proposal and the one that decides whether the thing is still working in March.

Here are real numbers for all three.

The token bill is the part everyone worries about and the part that matters least

Start here, because it is the number people ask for and because seeing it usually reframes the whole conversation.

Claude is billed per million tokens, in US dollars. The current rate card:

• Claude Haiku 4.5 — $1 input / $5 output per million tokens
• Claude Sonnet 5 — $2 input / $10 output
• Claude Opus 5 — $5 input / $25 output

Now a worked example. The token volumes below are our own illustrative assumptions, not vendor figures — your real numbers depend entirely on how much context your agent carries.

Say a customer-support agent handles 10,000 conversations a month, and each conversation costs roughly 3,000 input tokens (the system prompt, the retrieved knowledge, the conversation so far) and 700 output tokens of reply.

That is 30 million input tokens and 7 million output tokens. On Sonnet 5:

• Input: 30 × $2 = $60
• Output: 7 × $10 = $70
• Total: about $130 a month — roughly RM525, at USD to MYR 4.04, the rate on 22 August 2026.

Ten thousand handled conversations for the price of a decent office chair, monthly. That is the number that surprises people, and it is why the token bill is rarely where a project succeeds or fails.

A caveat that matters more than the figure: Anthropic bills in US dollars. The rate card is stable; the Ringgit number is not. Quote it in USD internally and convert when you need to — and if you want your own figure rather than ours, our Claude cost calculator will work it out against your actual volumes and export it.

Three levers that genuinely move the bill, and one myth

If the token bill does matter at your volume, three things move it. They are worth understanding because the internet gets two of them wrong.

Right-size the model per task. The same workload above on Haiku 4.5 costs $30 input plus $35 output — about $65 a month, roughly RM263. Half the price. Not every task needs the strongest model, and defaulting everything to the most capable one is the most common source of overspend. Classification, routing, extraction and short replies rarely justify Opus.

Prompt caching, with the detail almost everyone states backwards. Cache hits cost 10% of base input. Writing to the cache costs more than a normal call: 1.25× base input for the five-minute TTL, 2× for the one-hour TTL.

The break-even is where the confusion lives. On the five-minute cache, a single read already pays for the write. On the one-hour cache, you need two reads before you are ahead. A lot of published advice says "caching only pays after several reads" without naming a TTL, and on the five-minute duration that is simply wrong. The only unambiguous loss is a cache write nobody ever reads.

Applied to the example above: if 2,000 of those 3,000 input tokens are a stable system prompt and knowledge base that can be cached, the input cost falls from $60 to roughly $24 — the cached portion billing at $0.20 per million instead of $2. Total drops to about $94 a month, near RM380. Cache writes are real and I have left them out of that figure for clarity; at ten thousand conversations they amortise to very little, but they are not zero.

The Batch API is 50% off input and output for work that does not need an immediate answer — overnight document processing, bulk classification, report generation.

And the myth: caching and batching do not stack on the same workload. Batch is for asynchronous jobs; caching pays on live, repetitive traffic. Anyone presenting them as two multiplying discounts on one workload has not run either.

One more thing, because it is counterintuitive and nobody writes about it. Sonnet 5 uses a newer tokenizer that produces approximately 30% more tokens for the same text. That sounds like a hidden price rise. It is not: Sonnet 5 sits at $2 against Sonnet 4.6's $3, so the same text works out cheaper on the newer, better model even after the token inflation. If someone quotes you a cost comparison between Claude versions without accounting for the tokenizer, the comparison is wrong. We go deeper on plan and API pricing in our Claude pricing guide for Malaysia.

What actually drives the build cost

This is the large number, and it is not driven by the model. It is driven by how much of your business the agent has to touch.

The integration surface is the single biggest factor. An agent that answers questions from a document you upload is a modest piece of work. An agent that reads your order system, checks stock, issues a refund and writes the result back to your accounting package is four integrations, four sets of credentials, four failure modes and four vendors whose APIs behave differently on a bad day. Cost scales with connections, not with cleverness.

Approval design is the second. Deciding what the agent may do alone, what needs a human to approve, and what it must never attempt is genuinely hard work, and it is where the risk lives. An agent that can issue refunds without a ceiling is a very expensive agent to have built cheaply.

Exception handling is the third, and the most underestimated. The happy path is perhaps a fifth of the build. What happens when the customer's order number does not exist, when the API times out halfway through, when the request is in Bahasa Malaysia and the knowledge base is in English, when somebody tries to talk it into something it should not do? Every one of those is a decision that has to be made and coded.

The pattern to watch for in quotations: a low build price usually means the exception paths were not scoped. You do not save that money. You spend it later, in production, with your customers finding the gaps.

The cost nobody quotes: keeping it running

An AI agent is not a website. A website you build, launch and leave alone. An agent sits in a changing world.

The model gets updated. Your product catalogue changes. A supplier alters their invoice format. A customer phrases something in a way nobody anticipated. An agent that was 94% accurate at launch can be quietly making a specific kind of mistake by week seven, and unless somebody is looking, the way you find out is from a customer.

So a real running budget has four items, and if your quotation contains none of them, that is the finding:

• Monitoring — what is measured, how often, and what threshold triggers a look.
• An exception queue with a human in it — the cases the agent declined or got wrong, reviewed by someone whose job includes it.
• Periodic review — a scheduled pass over a sample of real interactions, not an alert that only fires when something breaks loudly.
• Named ownership — a person or rotation. "The team" is not an owner.

None of this is glamorous and all of it is where the ongoing cost of AI actually sits. It is routinely larger than the token bill and routinely absent from the proposal. It is also question two of the seven we suggest asking in how to choose an AI agency in Malaysia.

When an agent is the wrong purchase

Worth saying plainly, because we would rather tell you now.

If the task runs a handful of times a week, a person doing it is cheaper than a system that has to be built, monitored and maintained. If the process is not written down anywhere, an agent will automate whatever it guesses, and you will have expensively encoded a misunderstanding. If the answer changes depending on judgement nobody can articulate, that judgement cannot be specified, and the agent will be confidently wrong at scale.

The tasks worth automating are high-volume, well-defined, and currently done consistently by a human who could explain the rules. That is a narrower set than most vendors imply, and it is a very profitable set when you pick from it correctly.

So what should you budget

Honestly: the token bill for a well-scoped agent is usually tens of Ringgit to a few hundred a month, and you can work yours out to the cent with the calculator.

For the build, we publish our range rather than making you ask. Scoped agent projects at Anchor Sprint run from RM5,000 for a single automated workflow to RM50,000 for a deployment spanning several systems with approval steps and audit requirements. The spread is governance, not intelligence — the expensive end is expensive because of how many systems it touches and how carefully it has to fail, not because the model is cleverer.

The running cost is a real monthly line item that most quotations omit entirely.

If a proposal shows you a large build number, a token estimate, and nothing about who watches it afterwards, you are looking at two of the three costs.