What We Could Not Confirm, and Why That Matters

There are entries on this site that say, plainly, "not published." For most of this site's life the clearest examples were the Cyber-named models: no price, no context window, no rate limit — just a name and an honest admission that nothing checkable existed for them. There's a temptation to read a gap like that as a failure of research, something to apologize for or quietly hide. This site treats it as the opposite: the clearest evidence a reader has that everything else on the page was actually checked.

Why "not published" is a claim, not a shrug

It would be easy to treat "not published" as the lazy default — the thing you write when you didn't look hard enough to find an answer. That's exactly backwards, and it's exactly the mistake this site's own research has made repeatedly before correcting course. Writing "not published" honestly requires having checked the places a fact would plausibly live and coming back empty, not just failing to stumble across it. A model's own reference page, the general pricing page, the rate-limits guide — checking all of the relevant surfaces, and naming which ones were checked, is what turns "I don't see it" into a defensible "it isn't published."

What happened to the models that carried this label

When this site launched, GPT-5.4 Cyber and GPT-5.5 Cyber appeared in the pricing page's source with every price cell empty, and their model pages returned nothing. By this autumn's re-check they had gone in opposite directions. GPT-5.5 Cyber gained a Standard price row in the pricing page's Cyber table — but its model page still returns a 404, so its window, cutoff and rate limits remain unpublished, and this site shows the price alone rather than borrowing specs from a sibling. GPT-5.4 Cyber was deprecated and then removed from the API without ever having had a published price. Meanwhile a newer model, GPT-5.6 Cyber, arrived with a full model page and price row, and it requires separate approval through OpenAI's Daybreak program.

That sequence is the argument for the label in miniature. A gap closed with a real number when OpenAI published one; a gap that never closed ended with the model's retirement, not with a guess that would now be quietly wrong. None of the Cyber models is offered in any calculator here — the unpriced one because there's nothing to compute with, the priced ones because they're approval-gated, and a cost estimate for a model most readers can't call isn't an answer to anything. The model retirement schedule tracks the rest of the lineup's exits.

Why guessing would be worse than the gap

Imagine the alternative: a calculator that, faced with an unpriced model, quietly substituted a nearby sibling's rate rather than refusing to offer it. The output would look complete. It would also be confidently wrong, in a way a reader would have no way to detect, because nothing on the page would signal that the number was invented rather than sourced. A visible gap is honest. An invisible substitution is a lie dressed up as data.

The corrections that taught this site the opposite failure mode

It's worth being equally honest about the other direction, because this site has made that mistake too: over-using "not published" for a figure that actually was published, just not where it looked. Per-model rate limits were reported unpublished after checking only a general guide that defers to account settings, when each model's own reference page carried the real table. The long-context repricing rule was reported unpublished after checking only the pricing page, when the model's own page states it in a plain sentence.

The most recent correction is the most instructive, because the right page was checked. This site reported that Sol models published no reasoning-effort levels, when the levels were on those pages all along — written in a sentence shaped differently from the one its scraper had been built to recognize. The scraper's failure to parse the sentence was recorded as the page's failure to say anything. It's the same mistake as the others, one layer down: absence of a successful parse reported as absence of the fact. The fix was structural, not a patch. The scraper now reports anything it can't parse as unparsed, never as unpublished, and stops the run until a person looks; and a build check compares every model figure on this site against freshly scraped evidence, so a transcription slip can't ship either. Every one of these corrections mattered as much as the original gap would have — a false "not published" quietly removes a real, useful figure from every calculator that could have used it.

A different failure: prose that drifted from its own facts

The same re-check turned up a quieter problem. Some pages' prose, written at launch, contradicted this site's own fact tables — a claim about a model's features or limits that had been true of an earlier reading, or never quite true, and had never been re-read against the data behind it. No figure was wrong in the tables; the sentences around them were. Those pages were rewritten, and prose here now pulls its figures from the same facts module as the tables rather than restating them, so a corrected fact corrects every sentence that mentions it.

The discipline this creates

Every fact entry on this site records which specific pages were actually opened, not just a final verdict. That's a small, unglamorous piece of bookkeeping, and it's the entire mechanism that makes the difference between an honest gap and a lazy one auditable rather than a matter of trust. A reader who wants to check whether "not published" is well-founded can see exactly what was checked and decide for themselves whether that's sufficient — see how this site sources every number for the full process.

Why this matters more here than on a lot of sites

A site organized around computed answers — prices, break-evens, budget projections — lives or dies on whether its inputs are real. A single invented number, even a plausible-looking one, corrupts every downstream calculation that touches it, silently, with no error message telling the reader something's wrong. Treating "not published" as a real, checkable, sourced answer — as legitimate an outcome as a number — is what keeps that corruption from happening quietly. It's a narrower promise than "here's everything you need to know," and it's the right amount of confidence to have about facts a vendor genuinely hasn't published.

What to do if you find a gap that's since been filled

Vendor documentation changes — the Cyber models above are proof. A model that's unpublished today may not be next month, and this site's weekly re-verification exists to catch that kind of change rather than letting a "not published" label go stale into a false one. If you've found a published figure this site still lists as unconfirmed, that's worth reporting: closing an honest gap with a real number is exactly the kind of correction this site's process is built to make.

Why "unpublished" is not the same claim as "doesn't exist"

It's worth being precise about what this label does and doesn't assert. "Not published" says nothing about whether a model, a feature or a mechanic exists internally at OpenAI — only that no checked, citable source states it publicly. A reader planning around one of these gaps shouldn't read it as "this will never be priced" any more than as "this is definitely coming soon." GPT-5.4 Cyber's history shows both readings would have been guesses: it was never priced, and it no longer exists. The label is a statement about the current state of public information, with no implied prediction about what happens next.

How this compares to simply omitting the entry

An alternative approach would be to leave unpriced models off this site entirely — no page, no mention, nothing for a reader to find. That would avoid ever publishing an unconfirmed figure, but it would strand exactly the reader who most needs an answer: someone who's met one of these model names somewhere and is trying to find out what it actually costs. A page that says plainly "here's what we checked, and here's what we found: nothing" serves that reader far better than silence does, even though silence would look, superficially, like the more cautious choice.

The standard this sets for every other claim on the site

Once a site is willing to publish "we don't know, and here's exactly what we checked," it raises the bar for everything else it claims to know — a positive figure sitting next to an honest gap reads as more credible than the same figure would on a site that never admits a limit to what it's confirmed. That's not an accident of presentation; it's the mechanism by which admitting uncertainty in the right places makes the certainty elsewhere more trustworthy, not less.