What does it cost to add AI to an app?
Two numbers matter and they behave completely differently. The build cost is a one-off — typically AED 18,000 to AED 40,000 for a production-grade AI feature including prompt engineering, evaluation, guardrails and fallback handling. The running cost is per request, forever, and scales directly with how many people use it.
Teams routinely get the first number right and the second one badly wrong. A chatbot that costs $40 a month in testing can cost $4,000 a month at 100,000 users — and nobody modelled it, because in development the bill was a rounding error.
How LLM pricing actually works
You pay per token, separately for input and output, and output is roughly five times more expensive than input. A token is about four characters of English, so 1,000 tokens is roughly 750 words.
Current Anthropic list prices, per million tokens:
- Haiku 4.5 — $1 input / $5 output. Fast and cheap; ideal for classification, extraction, routing and short replies.
- Sonnet 5 — $3 input / $15 output. The sensible default for most product features.
- Opus 5 — $5 input / $25 output. Worth it for genuinely hard reasoning, long agentic work and complex analysis.
The input side is where volume hides. A RAG feature that retrieves six documents per question sends thousands of input tokens per request even when the answer is two sentences long.
Prompt caching is the biggest lever you have
Most AI features send the same large prefix every single time — a system prompt, a tool list, a knowledge base excerpt. Prompt caching lets the provider reuse that prefix at roughly one tenth of the input price, and it is usually a few lines of code.
Toggle it in the calculator to see the difference. On a document-heavy feature it commonly removes half the monthly bill. The catch is that caching is a prefix match, so the stable content has to come first and anything that changes per request has to come last — which is an architecture decision, not a config flag.
What the build cost actually covers
Wiring an API call to a text box takes an afternoon. Shipping an AI feature to production takes considerably longer, and the difference is the part users notice:
- Evaluation. A test set of real questions with known-good answers, so you can tell whether a prompt change made things better or worse.
- Guardrails. Handling refusals, off-topic requests, prompt injection and the cases where the model should hand off to a human.
- Retrieval. For anything grounded in your own data: chunking, embedding, vector storage and keeping the index fresh.
- Cost controls. Rate limits, per-user caps, caching and model routing — send the easy 80% to a cheaper model.
We cover the approach in more depth on AI automation services and AI chatbot development in Dubai.