Inference Cost Model — Reliova
Spec · Free tool

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.

00 · What you run today
Pick “Nothing yet” if the feature has not launched — we will compare against the usual default choice instead.
01 · Workload shape
02 · Scale
03 · Shape parameters
04 · Execution
Stable system prompt and history re-sent each turn
Applies to reasoning models only
Closed frontier — cheapest
price only
Open-weight — cheapest
i These are the cheapest option in each group, compared on price alone — they are not matched on capability. Whether a cheaper model can actually carry your task is an evaluation question, not a pricing one, and it is the part the free report answers.
Estimated monthly cost by model Open-weight Closed frontier
Model Eff. in Eff. out Monthly (p50) p90 Per 1k users
Where the tokens go
Throughput feasibility
What is an estimate, and what is a commitment
This page An estimate. Not a promise. It runs on numbers you typed and rates you can edit, and it compares price alone. We will not commit to a saving figure before seeing a single one of your real prompts — anyone who does is guessing with your budget.
The rate Anyone can buy at a published rate. A rate does not tell you which model actually carries your task, whether the provider is still serving it at full precision next month, or who answers the phone when it stops. Those three are the job — the rate is just the invoice.
Free report A finding on your real traffic. We build an evaluation set from your own production samples and report, per task type, what moves without measurable quality loss and what does not. Free, and yours to keep whoever you end up buying from.
Scoped against the report. The work is quoted against what the report found — which task types move, at what measured quality — so the scope is settled before anyone starts. Commercial terms are agreed in the engagement, not assumed from a web page.
Priced to the commitment, and watched after it starts. Committed volume is quoted as a term rate rather than a moving list price. We fingerprint what is actually served, so a drift away from the agreed model or precision shows up as evidence rather than a suspicion. How that is handled is written into the agreement.
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:

input(T) = T·S + T·u + (u + a) · T−1/2 output(T) = T·a

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:

cacheable(T) = T·S + (u + a) · T−1/2 cached = cacheable × hitRate fresh = input − cached

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:

input(S) = [ S·B + (o + r) · S(S−1)/2 ] · (1 + ρ) output(S) = S · o · (1 + ρ) · thinkingMultiplier

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:

effectiveRate = offPeak × (1 + peakShare) US business hours peakShare ≈ 0.00 Europe business hours peakShare ≈ 0.35 Global, round-the-clock peakShare ≈ 0.21 Asia business hours peakShare ≈ 0.78

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:

requiredTPM = monthlyTokens / activeMinutes × 3 requiredRPM = monthlyRequests / activeMinutes × 3

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
! These are public market rates — they are not Reliova rates. They are here so the model has a neutral, checkable baseline. Reliova quotes against committed volume, so our rate is given in a quote, not published on a page. Type your own quoted rate into any row and the model re-runs on it.

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.

Get the free report →