The tool stack and what it costs in PKR

Three ways to see it

1. 1

   The model layer. For most Pakistani teams in 2026, the practical choices are Anthropic Claude (Opus, Sonnet, Haiku), OpenAI GPT (4, 4o, o1), and Google Gemini (Pro, Flash). All three publish per-token pricing in USD. A rough current floor at PKR 280 per USD: Claude Sonnet sits around PKR 1 per thousand input tokens, PKR 4 per thousand output tokens; smaller models like Haiku and 4o-mini are a tenth of that. A practical Urdu sentence is roughly 30 tokens. A typical FBR-return summary prompt is 5,000 input tokens and 2,000 output tokens, giving a per-call cost in the PKR 10 to PKR 15 range with mid-tier models. Sixteen returns a day means PKR 5,000 to 7,000 per month in model costs. That number changes by an order of magnitude depending on which model you choose and whether you cache.

2. 2

   The orchestration layer. You need a small framework to chain prompts, manage tools, and store context. Anthropic and OpenAI SDKs are free; LangChain and LlamaIndex are open source; commercial wrappers like Vercel AI SDK, n8n, or Make.com sit on top. Most Pakistani teams should not pay for an orchestration platform in year one. A junior engineer in Python or TypeScript can build the orchestrator in a week. The total layer cost is zero, the cost is in engineering time. Plan for two weeks of one engineer's salary, somewhere in the PKR 100,000 to 250,000 range depending on seniority.

3. 3

   The data layer. Your corpus needs storage and a search index. For small pilots, plain Postgres with the pgvector extension or SQLite with FTS5 is enough; cost is the cost of one virtual machine, around PKR 10,000 a month. For larger pilots, managed vector databases like Pinecone, Weaviate Cloud, or Qdrant Cloud charge USD 50 to 500 per month for the free-to-low-mid tiers. The hidden cost is data preparation: cleaning, chunking, embedding, and re-embedding when documents change. Budget engineering time for this honestly. The cleaner your corpus, the cheaper everything downstream.

The why-tree

Try this with Claude

AI-edge prompt to try with Claude or ChatGPT: 'I am running a pilot at a Pakistani bank: an LLM that summarises 200 customer complaint emails per day with citations. Build the bill of materials in PKR over six months. Assume 1 USD = PKR 280. Compare three model choices (Claude Sonnet, GPT-4o, Gemini Pro), show me where I save by caching, and call out the biggest hidden cost.' Cross-check the numbers against vendor pages.