Tool

Fully loaded inference cost estimator

How far will the bill exceed the headline rate? Runs in your browser. Nothing is stored or transmitted.

Build a monthly inference estimate from the rate card up, adding the mechanics a headline rate leaves out. Showing illustrative defaults. Replace them with your provider's current terms and pilot data.

Volume and shape

Requests a month

Calls to the model, not users. An agent task can make many.

Input tokens per request

Prompt, instructions, retrieved documents and conversation history.

Output tokens per request

Answer plus any reasoning tokens. Measure in a pilot.

Rate card, per million tokens

Input price

From the provider's current price list.

Output price

Often four to six times the input price.

Caching

Input that is reusable

Stable system prompts and shared documents.

Cache hit rate

Share of reusable input actually served from cache. Falls with bursty traffic.

Cache read price, as % of input price

Commonly a tenth of the input price, and as low as 2.5% on some models.

Cache write premium above input price

Some providers charge extra to write to the cache.

Side-charges

Uplift on token cost

Regional processing, long-context tiers and similar multipliers. Some models charge 100% to 400% more once a prompt passes a size threshold, such as 100,000 tokens.

Other fees per request

Tool calls, web search, retrieval and guardrail checks.

Fully loaded a month

£5,025

Per request

£0.025

Against headline

1.5×

Headline £3,264

The fully loaded bill is £1,761 a month above the headline estimate. The largest single adjustment is output premium, at +£1,536.

From headline to fully loaded

Headline: every token at the input price£3,264
Output premium+£1,536
Caching saving−£959
Uplifts on token cost+£384
Tool, retrieval and guardrail fees+£800
Fully loaded£5,025

The headline figure prices every token at the input rate, the way a quick comparison of rate cards often does. The estimate leaves out rate-limit tiers, which add friction and lead time rather than a charge, and any platform, people or evaluation cost outside the model invoice. Put those in the AI TCO worksheet.

The four layers are explained in The Real Cost of AI Inference. To test the estimate against higher prices or demand, use the AI budget stress test. If volume is high and steady, the self-hosting break-even calculator shows whether dedicated capacity would cost less.