TRANSPARENT BY DESIGN

Understand every number.

The calculation methodology behind AI Cost. Price snapshot checked on October 12, 2026.

Monthly API cost

We price each request class separately. For each model, total input is uncached input + cache-read input + cache-write input. These are non-overlapping billing categories: do not put a cached token in uncached input again.

Choose the short or long tier using the complete input length for that class. Apply its rates to all input categories and billed output, add the explicit per-request extra, then multiply by daily requests and your selected days per month.

request cost = (input × input rate + cache read × read rate + cache write × write rate + output × output rate) / 1,000,000 + extras

Open the monthly calculator →

Long inputs and cache writes

Claude Haiku 5.5 changes tiers above 100,000 input tokens. GPT-6 Luna changes above 272,000. Exactly at the threshold still uses the short tier. We never price only the excess tokens at the higher rate.

The default is no cache. If you have cached usage, enter distinct read and write quantities. Haiku supports the shown 5-minute and 1-hour cache-write rates; Luna uses its published Standard cache-write rate. Enter billable usage from your provider, not an assumed cache-hit percentage.

A custom rate replaces both official tiers with the flat USD rates you enter. The result is labeled user-defined. Model context and output constraints remain in place; this is not a simulator for a different model.

Review the two-model price table →

Cost per accepted task

T is the number of business tasks started. A is the total billed API calls including retries and multiple steps. S is the number of tasks finally accepted under one shared standard, with each task counted at most once. C is the entire sample cost, including unsuccessful calls and any costs you choose to include.

Enter C from a representative sample or invoice calculation. We do not multiply that amount by retries again. A / T describes the observed calls per task; it is not a second cost multiplier. The current sample workflow assumes every started task made at least one billed API call.

When S = 0, there is no finite observed cost per accepted task. Without a completed sample, the tool makes no quality claim. If the included sample cost is zero, a fixed budget cannot establish a finite task capacity.

cost per started task = C / T
cost per accepted task = C / S
monthly spend = monthly started tasks × C / T
monthly accepted tasks = monthly started tasks × S / T

Learn how to define accepted tasks →

Quality, latency, and budget

Eligibility requires at least one accepted task and your minimum final acceptance rate. If you set a maximum average latency, each model also needs your measured latency at or below that limit. Missing latency is not treated as passing.

Budget capacity is floor(budget / observed cost per accepted task). Monthly accepted volume is an expected value, so it may be fractional. These projections assume a similar task mix and workflow; they are not a guarantee of future quality, spend, or throughput.

What is not automatically included

Batch, Flex, Fast, regional processing premiums, server-side tools, image or audio billing, hosting, retrieval, storage, taxes, discounts, and account credits are not automatically calculated. Monthly estimates are resource costs before credits. Add known per-call charges explicitly, or include all relevant charges in your observed sample total.

The calculator does not tokenize prompts or run a model. It cannot reconstruct an invoice from a monthly token total that mixes different per-request tiers. Use billed token counts, including billed reasoning output, from your own logs.

Price maintenance and data handling

Official rates are a dated snapshot, stored once and reused in every view. A warning appears when that snapshot is more than 30 days old. Saved plans reload against the current price snapshot; if the saved version differs, the tool tells you.

Calculations run in the browser. Local saving is opt-in. Share links contain only the allowlisted numeric settings and calculator switches in the URL fragment; they do not contain prompts, API keys, or logs. Anyone holding the link can see its assumptions.

Plausible page analytics is configured separately from the calculator. We do not send calculator values as analytics events. There is no model API integration, account system, or payment processing.

Read the privacy policy →

Primary pricing sources

Found a pricing mismatch? Email hi@sevu.net with the official source and request conditions.