In this guide · 8 sections
Short answer
At October 2026 list prices, a short FAQ reply costs about US$0.0003 on a small model and US$0.006 on a mid-range model, before caching. A long answer drawing on documents costs roughly four times that. On WhatsApp, Meta's S$0.0205 service message fee (after 1,000 free a month) often exceeds the AI cost.
How AI reply costs are calculated
AI models charge by the token, a chunk of text roughly the size of a short word or part of a longer one. Every reply is billed on two counts:
- Input tokens: everything sent to the model: your instructions (often called the system prompt), any FAQ or document text you add, the conversation so far and the customer's question.
- Output tokens: the reply the model writes, plus any hidden reasoning tokens on models that think before answering.
The formula is simple: input tokens x input price + output tokens x output price, with prices quoted per million tokens. Output tokens cost about five times as much as input on the models below, but input is usually far larger, so both matter. Our AI API pricing comparison covers the wider model list, including Google Gemini and DeepSeek.
The prices used in this guide (October 2026)
Standard list prices in US dollars per million tokens, from OpenAI's API pricing (short-context rates) and Anthropic's Claude pricing, checked on 19 October 2026. Approximate and before tax; vendors change prices often.
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6 Luna (OpenAI) | $0.10 | $0.01 | $0.50 |
| Claude Haiku 5.5 (Anthropic) | $0.10 | $0.01 | $0.50 |
| GPT-6.1 Sol (OpenAI) | $2.00 | $0.10 | $10.00 |
| Claude Sonnet 5.5 (Anthropic) | $2.00 | $0.10 | $10.00 |
| Claude Opus 5.5 (Anthropic) | $4.00 | $0.20 | $20.00 |
| GPT-6 Astra (OpenAI) | $10.00 | $1.00 | $50.00 |
Both vendors also charge a little extra to write to the cache the first time (US$2.50 per million instead of US$2.00 for the mid-range models), and Anthropic's listed cache prices assume a 5-minute cache lifetime. Batch processing, for work that can wait, is half price at both.
Worked example 1: a short FAQ reply
A customer asks your website assistant about opening hours or delivery fees. Each reply sends about 2,000 input tokens (1,200 of instructions, 600 of FAQ text and a 200-token question) and gets back about 200 output tokens, roughly a short paragraph.
| Model tier | Per reply | Per 1,000 replies | Per 1,000 with caching |
|---|---|---|---|
| Small (Luna, Haiku 5.5) | US$0.0003 | US$0.30 | US$0.14 |
| Mid-range (Sol, Sonnet 5.5) | US$0.006 | US$6.00 | US$2.58 |
| Opus 5.5 | US$0.012 | US$12.00 | US$5.16 |
| GPT-6 Astra | US$0.030 | US$30.00 | US$13.80 |
For this kind of question, a small model is usually enough. The difference between a small and mid-range model here is a few dollars per thousand replies, so the real decision is about answer quality, not cost.
Worked example 2: a long answer from your documents
A customer asks something that needs your policy documents, product manuals or price lists. The system first looks up the relevant passages (this is called retrieval-augmented generation, or RAG) and sends them with the question. Say 8,000 input tokens (1,200 of instructions, 6,000 of retrieved text, 800 of conversation history) and 600 output tokens.
| Model tier | Per reply | Per 1,000 replies | Per 1,000 with caching |
|---|---|---|---|
| Small (Luna, Haiku 5.5) | US$0.0011 | US$1.10 | US$0.99 |
| Mid-range (Sol, Sonnet 5.5) | US$0.022 | US$22.00 | US$19.72 |
| Opus 5.5 | US$0.044 | US$44.00 | US$39.44 |
| GPT-6 Astra | US$0.110 | US$110.00 | US$99.20 |
Two lessons from this table. First, caching helps much less when most of the input changes every time. Second, how much text you retrieve drives the bill: sending 6,000 tokens of documents when 2,000 would do triples the input cost. Our explainer on context windows covers why more context is not always better.
Adding the WhatsApp fee on top
If your AI assistant answers on WhatsApp through the WhatsApp Business Platform, Meta charges per delivered message as well. Meta's WhatsApp pricing page reflects updates effective 1 October 2026: replies inside the 24-hour window after a customer messages you are service messages, and each business phone number gets 1,000 free service messages a month, with charges from the 1,001st. Singapore is now its own market on Meta's rate card.
Meta's SGD rate card lists S$0.0205 per service message to a Singapore number (approximate, as of October 2026). Our WhatsApp Business API cost guide covers the other categories and provider fees.
How to keep the cost per reply down
Cost controls for an AI assistant
Check the token counts in your vendor's console after the first week. Real usage is often different from estimates, and the console shows input, output and cached tokens per call so you can redo the maths with your own numbers.
Do it yourself or get help?
If you have a developer, the formula and prices above are enough to estimate your running cost before you build. For most small businesses the model fees are a few dollars a month; the bigger costs are building the assistant well, keeping its knowledge current and checking its answers. Our guide to AI chatbots for websites covers those parts.
Our AI Website Assistant (S$790) answers customer questions from your own content, with caching and a sensible model choice set up from the start, on accounts in your name so you pay the AI vendor directly at list price.
Frequently asked questions
At October 2026 list prices, a short FAQ reply of about 2,000 input and 200 output tokens costs roughly US$0.0003 on a small model such as GPT-6 Luna or Claude Haiku 5.5, and about US$0.006 on a mid-range model such as GPT-6.1 Sol or Claude Sonnet 5.5, before caching. Longer document-based answers cost several times more.
Prompt caching lets the vendor store the repeated start of your prompt, such as instructions and FAQ text, and charge much less to reuse it. On current list prices a cached input token costs about a tenth or a twentieth of the normal rate. In our short-reply example, caching cut the cost per reply by more than half.
Each new reply sends the earlier messages back to the model as input so it can follow the conversation. Input therefore grows with every turn, and a five-turn chat can cost several times a single reply. Capping or summarising the history keeps the cost under control.
Yes, if you use the WhatsApp Business Platform. Since 1 October 2026, Meta charges for service messages sent within the 24-hour customer window after the first 1,000 free messages per business number each month. Meta's SGD rate card lists about S$0.0205 per service message to Singapore numbers.
Test before choosing. Run 20 real customer questions through a small and a mid-range model with the same instructions, score the answers blind, and pick the cheapest one that passes. For simple FAQ answers, small models are often enough; complex or sensitive questions may justify a mid-range model or a human hand-off.