Direct answer: the AI cost per task is the number of tokens one job uses, multiplied by the token price of that model. The formula: (input tokens × input price + output tokens × output price) ÷ 1,000,000. One customer service reply of 1,200 input tokens and 250 output tokens costs Rp 3.86 to Rp 430, depending on the model you pick.

Updated 11 September 2026. Every price here comes from the provider's own page on that date. The exchange rate I use: 1 USD = Rp 17,561.81, from open.er-api.com at 00:02 UTC on 11 September 2026. For comparison, the European Central Bank reference rate on 10 September 2026 was 1 USD = Rp 17,574.

Limit: model prices change without a long notice, and the rate moves daily. Recalculate before you sign a contract that uses these numbers.

The AI cost per task formula

An AI model charges per token, not per question. A token is a piece of a word. One English word is roughly 1 to 2 tokens, so 1,000 tokens is around 700 words.

Prices are always written per 1 million tokens, and the input price differs from the output price. Output is almost always dearer, usually 4 to 6 times the input price.

Cost of 1 task = (input tokens × input price + output tokens × output price) ÷ 1,000,000
Monthly cost = cost of 1 task × task count × retry factor + fixed cost

An example with Claude Haiku 4.5, priced at 1 USD input and 5 USD output per 1 million tokens:

(1,200 × 1 + 250 × 5) ÷ 1,000,000 = 0.00245 USD
0.00245 × 17,561.81 = Rp 43.03 per task
AI cost per task formula: input tokens times input price plus output tokens times output price divided by 1 million
The formula this article uses throughout. The example numbers use the Claude Haiku 4.5 price on 11 September 2026.

API prices per 1 million tokens, 11 September 2026

This table holds the standard price with no discount, short context. The rupiah column is my own calculation at the rate above, not an official rupiah price.

ModelInput (USD)Output (USD)Cost of 1 chat reply
Gemini 2.5 Flash-Lite0.100.40Rp 3.86
GPT-5.6 Luna0.201.20Rp 9.48
Gemini 3.8 Flash0.753.75Rp 32.27
Claude Haiku 4.51.005.00Rp 43.03
Claude Sonnet 52.0010.00Rp 86.05
GPT-5.6 Terra2.0012.00Rp 94.83
GPT-5.6 Sol4.0020.00Rp 172.11
Claude Opus 55.0025.00Rp 215.13
Claude Fable 5.110.0050.00Rp 430.26
GPT-6 Astra10.0050.00Rp 430.26

The last column uses the same task: 1,200 input tokens and 250 output tokens. The gap between the first row and the last row is 111 times. Model choice, not price negotiation, decides your bill.

Three notes that come straight from the documentation:

  • The Gemini 3.8 Flash price of 0.75 USD and 3.75 USD runs through 31 December 2026. From 1 January 2027 it becomes 1.50 USD and 7.50 USD, which is double.
  • OpenAI charges long context at a different rate. GPT-5.6 Sol moves to 8 USD input and 30 USD output once your conversation passes the short-context limit.
  • Claude 4.7 and later use a new tokenizer that produces around 30% more tokens for the same text. The price per token dropped, but the token count rises, so recalculate with your own text.

AI subscription prices in Indonesia, 11 September 2026

A subscription charges per person per month, not per task. The ChatGPT and Google prices below are the rupiah prices they show for Indonesia themselves. Claude publishes in dollars, so its rupiah column is my calculation.

PlanPrice per monthNote
ChatGPT GoRp 75,000More messages, this plan may include ads
ChatGPT PlusRp 349,000Reasoning models, Codex, deep research
ChatGPT ProFrom Rp 1,889,0005 times the Plus usage
Google AI PlusRp 75,0002 times more Gemini access, 400 GB storage
Google AI ProRp 309,0004 times more access, 5 TB storage
Google AI UltraFrom Rp 1,579,000Up to 20 times more access, 20 TB storage
Claude Pro17 USD annual, 20 USD monthlyAround Rp 298,551 to Rp 351,236
Claude MaxFrom 100 USDAround Rp 1,756,181, choose 5 times or 20 times Pro

The annual Claude Pro price is billed up front at 200 USD. None of these plans include applicable tax.

Worked example: 1,500 chat replies per month

This example is a simulation with sample data, not a client result. Use it as a shape and put your own numbers in.

Starting condition. An online shop receives 1,500 chats per month. One reply uses 1,200 input tokens, holding the system instruction, the last 5 messages, and the product data. The reply is 250 tokens.

Step 1. Read the cost of 1 task from the table above.

Step 2. Multiply by 1,500.

ModelCost of 1 task1,500 tasks per month
Gemini 2.5 Flash-LiteRp 3.86Rp 5,795
GPT-5.6 LunaRp 9.48Rp 14,225
Gemini 3.8 FlashRp 32.27Rp 48,405
Claude Haiku 4.5Rp 43.03Rp 64,540
Claude Sonnet 5Rp 86.05Rp 129,079
Claude Opus 5Rp 215.13Rp 322,698

Step 3. Add the retries. Assume 8% of tasks fail and run again. The Haiku 4.5 bill moves from Rp 64,540 to Rp 69,703.

Step 4. Add the fixed cost. A small server at Rp 120,000 per month brings the total to Rp 189,703.

Observable output. The real cost per task becomes Rp 126.47, almost 3 times the raw model number. That is the figure worth deciding with, not Rp 43.03.

Cost comparison of 1 chat reply across 5 AI models, from Rp 3.86 to Rp 215.13
The cost order of 1 chat reply of 1,200 input tokens and 250 output tokens. The bar length shows the order, the number on the right gives the exact value. Model prices on 11 September 2026, rate Rp 17,561.81.

Subscription or API, which costs less

This question often lands in the wrong room. A subscription and an APIAPIThe official door 2 systems use to exchange data, without anybody copying it by hand.Open the glossary sell different things.

A subscription gives 1 person access to a chat app. That person types, reads, and pastes the result. You cannot run a 24-hour automation on a subscription, because the terms restrict it and the quota follows 1 person.

An API gives your program access. The program runs with nobody watching, and the bill follows usage.

Compare the 2 only when a person could do the work either way, for example a staff member rewriting product descriptions:

OptionMonthly costEquivalent task count
ChatGPT Plus, 1 personRp 349,0003,680 tasks on GPT-5.6 Terra, or 36,801 tasks on GPT-5.6 Luna
Claude Pro monthly, 1 personAround Rp 351,2368,163 tasks on Claude Haiku 4.5

My working rule: under 3,000 tasks per month with a person doing the work, a subscription is simpler. Above that, or when the work must run without a person, use the API.

5 costs teams forget

  1. Retries. A failed task is still billed. Record your failure rate, then multiply the cost by it.
  2. Thinking tokens. On reasoning models, thinking tokens are billed as output. Google states this openly on its pricing page.
  3. A cache write costs more than a cache read. On Claude Opus 5 a 5-minute cache write costs 6.25 USD while reading it costs 0.50 USD. Caching pays only when the same prompt repeats.
  4. The batch discount is not automatic. OpenAI halves the price on the batch and flex paths, for example GPT-5.6 Sol drops from 4 USD to 2 USD input. You must change how you call it, not merely wait.
  5. Long context raises the rate. A conversation that keeps growing makes every later task dearer, even before the long-context rate applies.

I cover how to cut operational AI cost as a whole in cost per task for operational AI tools.

Checklist for calculating AI cost per task

  1. Define 1 task clearly. One reply, one summary, one draft. Owner: the process owner.
  2. Measure input and output tokens from 20 real examples, not from a guess. Evidence: 1 note sheet.
  3. Take the model price from the official page on that day. Write the date down.
  4. Write the exchange rate you used and its date.
  5. Add the failure rate and the retries.
  6. Add the fixed cost, then divide by the task count.
  7. Compare with the cost of a person doing the same task. Stop when the gap is thin.

Questions I get often

What does one AI chat reply cost? At 1,200 input tokens and 250 output tokens, it costs Rp 3.86 on Gemini 2.5 Flash-Lite up to Rp 215.13 on Claude Opus 5, at the 11 September 2026 prices.

Why is the output price higher than the input price? A model produces output one token at a time, and that work uses more computation than reading the input. The gap is usually 4 to 6 times.

Is 1 token the same as 1 word? No. One English word is roughly 1 to 2 tokens, and Indonesian text often needs more. Measure with your own text, because language and style change the number.

Can I run an automation on a subscription? No. A subscription gives 1 person access to a chat app, while an automation needs an API billed by usage.

Which model is cheapest for simple work? On the 11 September 2026 table, Gemini 2.5 Flash-Lite and GPT-5.6 Luna are the cheapest. Test the quality on your own work before you move the whole load.

Do the rupiah numbers here stay fixed? No. The rate moves every day and a provider can change a price. Recalculate with the formula above before you decide.

Limits and your next step

This calculation measures the model cost. It does not measure the cost of building the workflow, cleaning the data, or the time your team spends reviewing the output.

I promise no saving either. I promise numbers you can check, so your decision stands on today's data.

If you want to know which tasks are worth moving to AI and what they would cost in your business, that is the work we do in AI Workflow Audit. The result is a priority list with an effort and impact estimate.

Sources