How We Source Every Number on This Site

Every price, every context window, every rate limit on this site traces back to a specific, dated OpenAI page — never a training-data recollection, never a plausible-looking estimate, never a number borrowed from a competitor's documentation because it seemed close enough. That sounds like a simple rule to follow. In practice it's the rule this site has gotten wrong three separate times in one afternoon, and the story of getting it wrong is more instructive than the rule itself.

The rule, stated precisely

A figure ships on this site only if it was read directly out of raw page content from an OpenAI-controlled domain, by a process that can point to the exact page and the exact date it was checked. Where OpenAI genuinely doesn't publish something, the honest value is "not published" — never a plausible number filled in to avoid an awkward gap.

The failure that taught this site to distrust convenient summaries

The first research pass on this site's pricing data used an automated page-fetching tool that summarizes a page through a smaller model before handing back the result. That summary silently dropped two entire pricing columns — cache-write prices, and the entire long-context pricing band — and flattened four distinct service tiers into what looked like one clean table. Nothing about the output looked incomplete. It read as a tidy, confident summary of a pricing page, and a calculator built on it would have understated long-context agentic workloads by roughly half while looking completely healthy the entire time. The fix wasn't a better prompt to the summarizing tool — it was abandoning summarization for pricing data entirely and re-extracting every figure directly from raw page bytes instead.

The failure that taught this site to distrust a single checked page

Twice more, in the same research pass, this site's own notes stated a figure as "not published" — not because it genuinely wasn't published, but because the specific page checked happened not to state it, and that silence got generalized to "nowhere does OpenAI state this." Per-model rate limits were reported unpublished after checking only the general rate-limits guide, which defers to account settings — but each model's own dedicated reference page turned out to carry a full, detailed table the guide simply didn't repeat. The long-context repricing threshold was reported unpublished after checking only the pricing page, where the detail lives in a client-rendered element easy to miss — while the model's own page states the rule in a plain sentence.

What both failures have in common

Read them together and the shared mistake is visible: trusting one surface's silence as evidence for the whole site's silence. A vendor's documentation isn't one page — it's dozens, and different facts live on different pages for reasons that don't always track a reader's intuition about where they'd expect to find something. "I checked the obvious page and didn't see it" is a much weaker claim than "not published," and conflating the two produced three genuine, avoidable errors.

What changed in how this site researches now

Every fact entry carries a record of which specific surfaces were actually checked, not just a final verdict — so "not published" always comes with an honest account of where that conclusion came from, rather than reading as an unexplained blank. Pricing tables get parsed from raw page structure by a dedicated script, never summarized through an intermediate model. And "not published" is treated as a claim requiring the same rigor as any positive claim — it gets checked against a model's own dedicated page, not inferred from a general guide's silence.

Why this level of care is the whole point of the site, not overhead

A developer-tooling review site lives or dies on whether its numbers are trustworthy enough to plan real infrastructure spend against. A single wrong price, silently believed, propagates into every calculator this site publishes and every decision a reader makes using them. The corrections logged here aren't a confession of past failure kept around out of a sense of obligation — they're the clearest evidence this site can offer that its own numbers get checked, not just asserted, and that when they're wrong, the record says so rather than quietly disappearing.

What this means for you as a reader

Every number-bearing page on this site carries a note on where its figures came from and when they were last checked. If something here looks wrong, or if OpenAI has since changed a figure this site hasn't caught up to yet, that's worth reporting — this site's own standing practice is weekly re-verification specifically because a developer-tools reference that goes stale silently is worse than no reference at all.

Why raw extraction beats a convenient summary, generally

The lesson from the dropped pricing columns generalizes past this one incident: any tool that compresses a page for convenience — a summarizer, an automated scraper with a fixed schema, a human skimming quickly under time pressure — can drop exactly the detail that mattered most, and the output will still look complete because the compression happened silently. There's no error message when a summary drops a column; the summary simply doesn't mention it, and nothing distinguishes "this table has four columns and I'm showing you all of them" from "this table has six columns and I dropped two." Treating every convenience layer between raw source and published figure as a place errors can hide, and minimizing how many such layers sit between the two, is the practical discipline this incident left behind.

What re-verification actually looks like in practice

This isn't a one-time audit that happened once and gets referenced indefinitely afterward. A dedicated script re-fetches every model's own reference page on a standing schedule, and the build itself is configured to treat a figure that hasn't been re-checked within a set window as stale enough to block publishing rather than silently continuing to display it. That's a deliberately strict standard — a fact this site can't currently vouch for as recently checked doesn't get to keep appearing on the page as if it still were.

The broader principle behind all of this

None of these individual fixes matter as much as the general posture behind them: treat every published figure as something that could be wrong until it's been checked against a primary source, directly, recently, and specifically — not as something that's probably fine because it was right the last time someone looked. That posture is slower than assuming figures stay correct by default. It's also the only posture that's actually caught this site's own mistakes before a reader had to.