What your AI feature will actually cost
A price-per-token table cannot forecast a bill. Conversations grow quadratically, caching cuts input by most of it, reasoning models bill hidden tokens, and some providers charge by time of day. This model accounts for all four. Nothing is sent anywhere — it runs in your browser.
| Model | Eff. in | Eff. out | Monthly (p50) | p90 | Per 1k users |
|---|
The method — every formula used above
1 · Conversations cost more than turns × price
Every turn re-sends the whole history, so input grows with the square of turn count. With system prompt S, user message u and assistant reply a, a session of T turns sends:
A 12-turn conversation therefore costs roughly four times a naive "12 × one turn" estimate on the input side. This single term is the most common reason inference budgets are wrong.
2 · Prefix caching removes most of that growth
Everything before the current user message is a stable prefix and can be served from cache at roughly a tenth to a fifth of the input rate. Cacheable volume and the resulting split are:
Note that input − cacheable = T·u exactly: the current user message is never cacheable. A workload with a large stable system prompt and a high hit rate can land in a completely different price bracket from an identical one without caching.
3 · Agent loops accumulate the same way
With base prompt B, per-step output o, tool-result tokens r and S steps, plus a retry rate ρ:
4 · Reasoning models bill tokens you never see
Internal reasoning is billed at the output rate but never rendered. A multiplier of 3× on visible output is a common working assumption for reasoning-heavy configurations; heavier settings reach 8×. Only models marked as reasoning-capable in the rate table receive it.
5 · Time-of-day pricing
Some providers price by clock. DeepSeek's published peak window is 01:00–04:00 and 06:00–10:00 UTC on weekdays — 35 hours of a 168-hour week — at twice the off-peak rate. The share of your traffic landing inside that window depends entirely on where your users are:
Batch work is scheduled off-peak by definition, so it never carries the peak surcharge.
6 · The band, not the point
Output length and session depth vary far more than averages suggest, and cost is driven by the tail. The p90 figure re-runs the model with output length at 1.8× and session depth at 1.35×. Budget against p90, not p50.
7 · Throughput feasibility
Monthly totals hide the peak. Required throughput assumes traffic concentrates into the active window for your time zone, with a peak-to-average factor of 3 for user-facing paths:
These are compared against a typical enterprise allocation of 10–50M TPM and 2,000–3,000 RPM. Exceeding them is a capacity conversation, not a pricing one.
What this model does not do
It does not judge whether a given model is good enough for your task — that requires an evaluation on your own data, and no calculator can substitute for it. It does not include cache-write surcharges, embedding or retrieval costs, or egress. And it is only as good as the rates below, which move constantly.
Reference rates — edit any of them
Public list rates in USD per million tokens, as observed on 11 September 2026. Providers change these often and the lowest available rate for a given model varies by provider.
| Model | Input | Cached in | Output | Tier | Reasoning | Time-of-day |
|---|
This is the estimate. The report is the decision.
A number is not an answer. The free Spec report runs this model against your real prompts and traffic, adds a good-enough evaluation of which models can actually carry each task, and tells you plainly which workloads should not move. No charge, no commitment.