AI cost methodology

Every figure in the AI section comes from a published rate and a formula you can see. This page lists the sources, the assumptions you can change, and what the totals leave out.

Sources and updates

• Rates are read once a week from two public price feeds, OpenRouter's model list and the LiteLLM price file, for a hand-picked list of models. Each row links the vendor's own price page, which is the source of truth.
• A value a person has checked against a vendor page can be recorded as an override. An override wins over the feeds and carries the date it was checked and the page it came from.
• If the two feeds disagree by more than 1% the previous value is kept and the row is marked "Check". If a price moves by more than 30% in one update, or a model disappears from both feeds, the change is reviewed by a person before it is published.
• If a feed cannot be fetched, the previous data stays and nothing partial is published.
• A price a vendor does not publish is shown as "Not offered", never as zero.
• Prices were last read on 3 Nov 2026. A notice appears on every page when the newest check is more than 14 days old.

How a task cost is calculated

tokens in  = words in  x tokens per word
tokens out = words out x tokens per word x reasoning multiplier

input cost  = (tokens in x (1 - cached share) x input price
              + tokens in x cached share x cache read price) / 1,000,000
output cost = tokens out x output price / 1,000,000

run cost     = (input cost + output cost) x batch factor
monthly cost = run cost x runs per month

The batch factor is one minus the vendor's published batch discount when batch is on and the model has one; otherwise it is 1 and the row says batch is not offered. If a model has no published cache read price, cached input is priced at the normal rate and the row says caching is not offered. Where a vendor publishes a higher price above a context length, requests above it use that price.

Assumptions you can change

• Tokens per word. One ratio of about 1.33 tokens per English word is used for every model. Real token counts differ by model and by language, so treat totals as a guide to scale, not a quote.
• Chats. Input for each turn is the system prompt plus every earlier message and reply plus the new message. The default system prompt is 150 words.
• Cached share and batch. Both start off. Caching only helps when much of each prompt repeats.
• Reasoning multiplier. Starts at 1. Models that reason before answering bill those tokens as output, which can multiply the output total.
• Task defaults. The word counts and volumes for each task are starting points, not measurements.

Tiers

Models are grouped into small, mid and frontier by hand. The groups exist so that like is compared with like. They are not quality scores, and nothing on the site ranks models by quality.

What is left out

• Taxes, support plans and negotiated or volume discounts.
• Tool, search and file-storage charges, image and audio tokens, and fine-tuning.
• Engineering time, monitoring and anything else outside the per-token price.

What the site does not estimate

The site does not estimate what it costs a vendor to serve a token for closed models such as GPT, Claude or Gemini. Vendors do not publish it, so any figure would be a guess. The discounts page shows only the discounts vendors publish.

AI section · Cloud cost methodology