Editorial Policy

Every article starts from a question we have had to answer ourselves, or one that readers keep asking: how much it costs to summarize a long PDF, which model is cheaper for a high-volume classification job, how much text actually fits in a given context window. If a topic cannot be tied to a concrete decision a developer or founder has to make, we do not write it.

Each article also has to add something the provider's own documentation does not: a worked example, a side-by-side comparison, a cost breakdown at realistic volumes. We would rather publish one piece that answers a question completely than three that restate a pricing page.

Prices come from two places. The site fetches live per-token rates through our own pricing endpoint, which reads from OpenRouter; behind it sits a reviewed price table in the repository covering 20 models, last verified on October 20, 2026. Every entry carries its own last-updated date, and the table is the fallback whenever the live feed is unavailable. A live price that strays more than 25% from the reviewed table is rejected and the reviewed value is shown instead, and price changes a provider has already announced for a future date are recorded in advance so they switch over on that date.

That table is not written from memory. Each rate is checked by web research against the provider's official pricing page before it is committed, and rechecked when a provider announces a change. Token counts, context windows and currency conversion follow the same rules the calculators use, documented in full in our pricing methodology.

Every article starts from primary sources: the provider's official pricing page, model documentation and changelog, read on the day we write. Worked examples are computed with the same formulas the calculators use, and the figures quoted in the text are re-run by script before publication, so the prose and the calculator on the same page cannot drift apart.

We use AI tools in that process and say so plainly: they help draft, structure and translate the text between English and Portuguese. They do not decide what is published, and they do not vouch for a number. Every article is reviewed and signed by a named author before it goes live; in that review each price, token count and calculation is checked against the provider's official page and the site's own calculators, and anything that cannot be confirmed is cut rather than softened.

We publish two articles per week, and never more than two in a single day. That ceiling is deliberate. A high volume of thin pages is exactly what search engines treat as scaled content abuse, and it is not how a reference site earns trust — a fixed calendar with human review behind each piece matters more than raw output.

So far we have published 28 articles, each one as a matched English and Portuguese pair, with a minimum of 800 words per language. We do not ship a piece in one language and leave the other waiting.

Pricing corrections are the highest-priority fix on this site. If a rate, a calculation or a claim in an article looks wrong, email contato@lwforge.com with the page, the model and the value you expected. We check the report against the provider's official documentation and correct the page once it is confirmed — usually within a couple of days.

When a correction changes a figure a reader may have relied on, we update the article rather than quietly removing it, and we treat provider price changes as a reason to revisit already published pieces, not only to write new ones.