How we work
How we work
Fieldbook to AI takes prices and specs from OpenRouter's public listing every day and adds short editorial verdicts; the verdicts are opinions, not test results.
Costs here use the prices AI companies charge software developers, so they show relative cost, not your monthly bill. See what ChatGPT, Claude and Gemini charge you.
Where the numbers come from
- Prices, how much a model can read, what it can take in, and release dates: OpenRouter's public models list. Refreshed daily by an automated job.
- These numbers are developer (API) prices, not the monthly plans for apps like ChatGPT, Claude or Gemini. For what the apps charge, see What should I pay?
- Speed is not shown, because we have no verified source.
How costs for everyday jobs are worked out
We multiply a stated number of tokens by each model's list price. A token is a piece of text, roughly three-quarters of a word; we assume a page is about 500 words. These are rules of thumb, and real use varies.
- Write a 1,000-word essay: You send a 300-token request (about 225 words) and get back about 1,350 tokens (about 1,000 words). (Assumed 300 tokens in, 1,350 out.)
- Have a 100-message chat: 100 back-and-forth turns. Each of your messages is about 40 tokens (30 words) and each reply about 200 tokens (150 words). Every turn re-sends the whole conversation so far, which adds up to about 1.2 million tokens sent and 20,000 received. No caching discount is assumed. (Assumed 1,212,000 tokens in, 20,000 out.)
- Summarize a 50-page PDF: 50 pages at about 500 words a page is 25,000 words, or about 33,400 tokens in. The summary is about 500 words, or about 700 tokens out. (Assumed 33,400 tokens in, 700 out.)
"Reads about N pages at once" is the model's size limit converted with the same rule of thumb: tokens × 0.75 ÷ 500, rounded. It is a limit, not a promise of accuracy on a huge document. Where a listing has different prices for very long inputs or different days, we use the headline price.
Smartness ratings (1 to 5)
We take the Artificial Analysis Intelligence Index that OpenRouter reports, and band it: 50 or more is 5, 40 to 49.9 is 4, 30 to 39.9 is 3, 20 to 29.9 is 2, under 20 is 1. "Not rated yet" means no score is available. A low rating can simply mean a model is built for one narrow job. Source: Artificial Analysis.
How models and picks are chosen
We start from the 466 entries in the upstream feed, remove duplicates, free and batch variants, aliases and router entries, then hand-pick 32 notable models. Best and budget picks are editorial judgments from the maker's description, price and smartness rating. We have not run our own tests yet. The "free option" on each task page is the free plan of a major app (for example Claude Free or ChatGPT Free), taken from our plans data; free plans have limits and can change.
Apps versus developer models
Most people use apps such as ChatGPT, Claude or Gemini. We show which company makes each model, but we have not verified which model a given app plan uses. App subscription prices are on What should I pay?: US list prices read from each company's own pages (or press, where marked), checked on the date shown, with a confidence label on every price. Exact usage limits are not published by the providers, so we only repeat their wording.
Corrections and disclosure
If a number looks wrong, check the linked OpenRouter listing first. Spot a mistake? Email us. Every fix we make is recorded in the changelog.
Who pays, and what that changes
Fieldbook to AI is an independent guide and is not affiliated with any AI company. Nobody pays to be listed or ranked. We may add sponsors to the weekly email or the site. Sponsors are always labeled as sponsors, and they never choose or change our picks, prices or wording. Some links may become affiliate links; if so, they will be labeled too. Details are on the sponsor page.