A solar calculator does it for nothing. A hosted LLM lights up several thousand tokens to reach the same answer.
A solar-powered calculator adds 10 + 10 for nothing. Daylight hits the photovoltaic strip, a few logic gates flip, and the display reads 20. No per-use cost, no subscription, no network round-trip, no telemetry. Send the same question to a hosted large language model and the model is not really adding — it is generating a probable string of text one token at a time, and every token is metered. The table below is what 10 + 10 actually costs when the answer comes from a model.
| # | Step | Input tokens | Output tokens |
|---|---|---|---|
| 1 | Load the platform system prompt | ~2,500 | — |
| 2 | Load tool / skill / connector definitions | ~3,000 | — |
| 3 | Load conversation context (turn start) | ~80 | — |
| 4 | Tokenize the user prompt "what is 10 + 10" | 7 | — |
| 5 | Forward pass through every parameter of the model to produce the first output token | (replays the input above) | — |
| 6 | Emit internal reasoning / "thinking" tokens | — | ~120 |
| 7 | Generate the answer token: "20" | — | 1 |
| 8 | Detokenize, stream, persist, log to telemetry, bill | — | — |
| Total | ~5,587 | ~121 |
At a current frontier-tier rate of roughly $15 per million input tokens and $75 per million output tokens, that single sum costs about 9.3 cents. The calculator costs zero. Ten employees asking the model ten times an hour for an eight-hour day works out to about $74 a day to reproduce what a $4 piece of plastic from 1985 does silently on a windowsill. The model is doing extraordinary work elsewhere. Addition is not it.