AI scribes feel like magic, but every consultation burns real compute. The two meaningful line items are speech-to-text and note generation, and both are publicly priced.
Modern batch speech recognition lists at fractions of a US cent per audio minute — Deepgram's Nova-3 tier, for example, is published at $0.0043 per minute, with multilingual processing carrying a premium and providers billing the full audio length, silence included. Frontier language models suitable for clinical drafting, such as Anthropic's Claude Sonnet tier, list at $3 per million input tokens and $15 per million output tokens.
The itemised consult
Put together, a short 2-minute consultation costs on the order of ₹2 to process end to end; a 10-minute consultation lands around ₹6–7; a long 20-minute encounter roughly ₹11–12. Streaming transcription — the live-text experience — roughly doubles the speech line item, which is why most products transcribe after the recording stops.
Why this number matters to doctors
Unit cost is the quiet force behind every pricing page in the category. It is why "unlimited" plans carry fair-use clauses, why free tiers are capped, and why per-seat prices differ so sharply between markets. A scribe priced far below its compute cost is subsidised — and subsidies end. Doctors evaluating a tool for daily use are entitled to ask the vendor a simple question: does your price survive my busiest month?