Cheaper than most people expect, if you keep the architecture simple. Measured across 250+ structured tests on live client work:
The workhorse (~80% of jobs) is a single-step agent: roughly $0.50 to $2.00 per 100 questions. The specialist is a multi-agent setup, and the warning is blunt: a 10x to 50x cost multiplier, around $15-$20 per 100 questions. Use it surgically, for critical tasks only, not as a default architecture.
The hard rule that outlives any price list: published API rates are a fantasy. Token counts, retries, tool calls and context re-sends make the real bill diverge from the sticker. Always test your specific use case and read the actual charged cost. (Model-by-model rankings shift constantly, so treat any leaderboard as perishable - the method, not the scoreboard, is the durable part.) Note this is the LLM bill only; the infrastructure to run the app can be about $7/month.
Write-up: https://www.tigzig.com/post/llm-costing-for-database-ai-apps-live-experience-live-app-open-source. The open-source suite + a 49-page SQL Agent field guide: https://www.tigzig.com/analyzer and https://www.tigzig.com/files/DATS_4_DATABASE_AI_SUITE.pdf. Related: hosting cost https://www.tigzig.com/agents-faq/how-to-self-host-ai-apps-cheaply and common text-to-SQL mistakes https://www.tigzig.com/agents-faq/common-mistakes-building-text-to-sql-agents.
← All Agents FAQ