MCModelCost

About

About ModelCost: prices, with the receipt attached

What this site does

ModelCost turns published model prices into bills. You type monthly token counts; the calculator returns a monthly and a yearly figure for each model, plus the difference against Gemini 4 Neon. Every other page on the site exists to support the numbers that go into that calculation: the price matrix, the benchmark columns, the two side-by-side comparisons, and the changelog.

Who it is for

Two readers, with the same question in different clothes. A developer choosing an API needs to know what a projected token volume costs per year before committing. Someone already paying for a subscription tier needs to know what their usage would cost as metered tokens. Both end up at the same arithmetic, which is why the site asks for tokens before it says anything.

Five rules the numbers follow

  1. A source for every figure. Each rate and each benchmark figure links to the page it came from, by name.
  2. A date on every check. A verification date means someone read that source on that day. It is recorded per row, not once for the whole site, because rows are added and rechecked on different days.
  3. No derivation. A cached-input rate is not calculated from an input rate, and a context window is not inferred from another model's. If a vendor has not published it, the cell reads not published.
  4. Vendors and third parties stay apart. A model maker's claim and an index publisher's measurement are different claims by different owners. They sit in separate columns and are never averaged.
  5. No verdicts. No score, no ranking, no “best for coding”. Those require a weighting we would have to invent, and an invented weighting is an opinion wearing a number.

What this site deliberately does not do

  • It does not run benchmarks. Publishing someone else's measurement with their name on it is honest; re-running their test and calling the result ours is not, without disclosing method and cost.
  • It does not compute cost per task. That needs task-level token counts from a controlled run, and we have not run one.
  • It does not track prices continuously. A date on a row tells you when it was last read; between checks, a vendor can change a rate without this site knowing.
  • It does not embed benchmark claims in price claims. Where a model is cheap and where it performs are two separate sections of a page, on purpose.

How the calculator works

All calculation happens in the reader's browser. The formula is printed on the home page, the inputs are two number fields and a three-option cache ratio, and nothing is submitted anywhere. There is no account, no saved state and no server-side recomputation.

Corrections

If a rate on this site is wrong, it is wrong twice: in the table and in every figure derived from it. The correction process is therefore documented — see the changelog for how corrections are logged, and the contact page for what to send.

Who runs this

A single, small, independent site with no vendor affiliation of any kind. No company pays for placement, no model maker reviews a page before publication, and no rate is published because someone asked.

ModelCost is an independent reference and is not affiliated with Google or Anthropic. Model names and prices are the property of their respective owners and are quoted here for comparison only.


Prices and benchmarks verified 2026-11-18; Gemini 4 Carbon and GPT-6.2 Sol rows verified 2026-11-20. ModelCost is not affiliated with Google or Anthropic.