Why model three tiers instead of one number?

Costs that look trivial at 10K requests a day become a major budget line at 1M. Seeing all three tiers side by side reveals where your margin breaks and forces a caching, routing, or pricing decision before you scale into it.

How is per-request cost calculated?

It multiplies your prompt tokens by the model's input price and completion tokens by its output price, both per million tokens, then sums them. That per-request cost is multiplied across each daily volume tier and extended to monthly and annual figures.

What does the cost-to-revenue ratio tell me?

It is your AI cost divided by revenue per request. Below roughly 30 percent you usually have healthy margin; above 100 percent each request loses money. It is the fastest sanity check that your unit economics survive scale.

Where does pricing come from?

The tool uses representative public per-million-token list prices. Vendors change prices and offer volume or batch discounts, so confirm against the provider's current pricing page before committing budget.

What is the At-Scale AI Cost Planner?

Projects daily, monthly, and annual LLM costs at three scale tiers for your chosen model and token profile, with a cost-to-revenue ratio input so you can see your gross margin at each level of volume. It runs free in your browser on Gera Tools, with nothing uploaded.

At-Scale AI Cost Planner

Name: At-Scale AI Cost Planner
Creator: Gera Tools
License: https://creativecommons.org/licenses/by/4.0/

At-scale AI cost planner

A model that costs a few dollars a day in a prototype can cost six figures a year in production. This planner takes one request’s token profile and projects it across 10,000, 100,000, and 1,000,000 requests a day — or a custom volume — so you can see daily, monthly, and annual spend at each tier and, crucially, what fraction of your revenue it eats.

How it works

You enter prompt tokens, completion tokens, and a model. The tool computes per-request cost from the model’s input and output prices (per million tokens), then multiplies that across each scale tier and extends it to monthly (×30) and annual (×365) figures. Adding revenue per request layers on a cost-to-revenue ratio and gross margin, so the table tells you not just what AI costs but whether the business math holds at volume. Everything runs in your browser.

Why three tiers change the conversation

Founders routinely discover costs that seemed negligible at 10K requests/day represent a six-figure annual line item at 1M. The psychological effect of seeing all three tiers simultaneously — rather than extrapolating one number mentally — is significant. It forces the right questions early: Does the business model support this? Which requests could be served by a cheaper model? Where does caching eliminate redundant calls?

Worked example: a summarization API

Suppose each request sends 500 prompt tokens (document excerpt) and receives 200 completion tokens (the summary). At a hypothetical model priced at $3 per million input tokens and $15 per million output tokens:

Per-request cost: (500 × $3 / 1,000,000) + (200 × $15 / 1,000,000) = $0.0015 + $0.003 = $0.0045
At 10K requests/day: $45/day → $1,350/month → roughly $16,400/year
At 100K requests/day: $450/day → $13,500/month → roughly $164,000/year
At 1M requests/day: $4,500/day → $135,000/month → roughly $1.64M/year

If you charge $0.01 per summary, the cost-to-revenue ratio at every tier is 45% — tight but workable. If you charge $0.005, you are losing money from day one.

Optimisation levers shown by the output

When the 1M-tier margin looks thin, the output immediately highlights which levers matter:

Lever	Effect
Prompt caching	Eliminates re-sending repeated system prompts; can halve input token cost for chat apps
Model routing	Cheap model for simple requests, expensive model only for hard ones
Output length control	Shorter completions cut the costliest token type — output is typically 3–5x the per-token price of input
Batch API discounts	Async batch jobs often qualify for 50% discounts on some providers
Price increase	A small price uplift at scale can recover more margin than engineering months of optimisation

Tips and notes

Use realistic token counts — measure a few real requests rather than guessing, since output tokens usually dominate cost and are easy to underestimate. If the 1M tier shows your margin collapsing, that is the signal to introduce a cheaper model for easy requests, prompt caching for repeated input, or a higher price. Batch APIs and volume discounts can cut the at-scale tiers materially, so treat list-price figures as the conservative ceiling. Always confirm live pricing before sizing a large deployment.