How we source every number on this site

Why sourcing discipline is the actual product here, not a footnote to it

A cost site whose numbers can't be trusted is worse than useless — it's actively harmful, since it looks exactly as authoritative as a trustworthy one right up until someone acts on a figure that turns out to be wrong. Given that, the sourcing process behind every figure on this site isn't a supporting detail; it's arguably the actual product, with the calculators and reference pages as the visible layer sitting on top of it.

The single reference every number traces back to

Every price, context window, cache minimum, and effort-related figure on this site is checked against one designated, in-repo reference treated as this site's factual spine, cross-referenced against the canonical public documentation it's built to reflect. Nothing is typed from memory or recalled from general familiarity with the space — a model id or a price recalled that way is exactly the kind of thing that's very likely stale by the time it's published, since this entire category of information changes fast enough that "familiar" and "current" are not the same property.

Why we don't treat "it sounds right" as a source

It would be faster, in the moment, to write a plausible-sounding figure based on general pattern familiarity with how pricing in this space tends to work, and faster is not the same as correct. A number that sounds right and is wrong is more dangerous than one that's obviously uncertain, because nothing about it prompts a second check. Every figure here traces to an actual, specific source, checked on an actual, specific date — not to a sense of what a plausible number in this range would look like.

What happens when a source doesn't fully answer the question

Sourcing isn't always clean. Sometimes the reference we treat as authoritative confirms a broader claim but doesn't restate every specific detail attached to it — a legacy model's exact context window, say, where its general specifications are documented but a specific figure isn't restated directly. In those cases, we don't fill the gap with an inference dressed up as a fact — we mark it explicitly as something we could not confirm, and the page renders that admission directly rather than papering over it with a plausible-looking number.

Why every fact carries a date, and why the date matters as much as the figure

A number without a verification date is a number you have no way to judge the trustworthiness of — it could be current, or it could be years stale, and nothing about the number itself tells you which. Every fact on this site carries the date it was last checked, and the site's own build fails once any of those dates goes too long without a genuine re-check, converting "we think this is still probably right" into an enforced, mechanical requirement rather than a hope.

The discipline that makes a re-verification honest rather than cosmetic

The easiest way to defeat this whole system is updating a verification date without actually re-checking the fact behind it — which produces a page that looks freshly confirmed and might not be. We treat that specifically as a worse failure than an honestly stale page, because it converts a visible warning sign into an invisible one. Every date on this site is supposed to follow a real re-read of the source, never precede it as a formality.

What this means for you as a reader

Trust the numbers on this site the way you'd trust any well-sourced reference: check the verification date if the figure matters for a real decision, and treat any explicit admission of uncertainty as more useful, not less, than a confident-looking number would have been in the same spot. That's the whole design, working as intended.

Why this process has to keep running, not just have run once

None of this sourcing discipline is a one-time setup task that, once complete, stays complete. The facts this site publishes describe a subject that keeps changing on its own schedule, and a sourcing process that ran correctly at launch and then stopped is functionally indistinguishable, a year later, from a site that never had one — both end up presenting confident, unverified numbers, just on different timelines. Treating this as a standing, recurring obligation rather than a completed milestone is the only version of this discipline that actually holds up over time.

See what we could not confirm, and why that matters for more on the admission side of this discipline, and any model's reference page to see the sourcing and dating applied directly, not just described.