GPT-6.5 Returns: A Hidden "6.5", a November 1 Countdown, and No OpenAI Confirmation

Six weeks ago this page reported that the name GPT-6.5 had died. The analyst who put it on O​penAI's next flagship retracted it within about twelve hours, and the model everyone had been calling "Doug" turned out — on his own corrected account — to be the base under GPT-6 Astra, which had already shipped on September 3. The label is back. This time there is an artifact behind it rather than a codename: a small site at opensi.lol whose source carries a hidden element identified as object-6-5, a "6.5" no visitor is meant to see, and a countdown that ends on November 1, 2026 at 10:00 AM San Francisco time. There is still no O​penAI statement, no model identifier, no picker entry and no catalog listing, and the six weeks in between were not idle ones: O​penAI cancelled a planned G​PT-6.1 Astra release over safety findings and shipped GPT-6.1 Sol in its place at a fifth of Astra's price. So the job here is the same as it was in September, only harder: separate what can be checked from what is being inferred, because with this particular artifact the gap between the two is wider than the excitement suggests. The documented AI failures, retractions and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

What is actually on the page

The site is one screen. It is titled "OpenSI — Request a Reset," and its own meta description says the same. The visible furniture is a counter labelled "requests worldwide," a button reading "Request a Reset," and an embedded post from O​penAI's Tibo Sottiaux dated October 19, 2026 — the one in which he says he bought a new domain and is "very excited for it." Pressing the button POSTs to an endpoint that increments the global count; when this piece was written the counter read 4,260,427 — a count of button presses rather than of visitors, but enough to show the page has been passed around hard in a single day.

The part that travelled is not visible at all. Below the fold, the markup contains an aside with the id object-6-5, marked hidden and aria-hidden — invisible to any reader, visible to anyone who opens view-source. Inside it are three things: a span reading "6.5," a coordinate pair reading 37.8067, -122.4314, and a time element whose datetime attribute is 2026-11-01T10:00:00-08:00, with fallback text counting down to "November 1, 2026, 10:00 AM San Francisco time." The coordinates resolve to Fort Mason Center on Marina Boulevard in San Francisco, an events venue. The time zone offset is -08:00 rather than -07:00, which is correct: daylight saving ends in the United States at 2 AM that same morning, so 10 AM on November 1 really is Pacific Standard Time.

That much is checkable, and this blog checked it directly rather than taking a screenshot's word for it: the page is live, the hidden block is in the served HTML, and the countdown attribute says what the posts say it says. What the page does not contain is any statement of who made it or what the date is for.

Why November 1 is the date that fits

The date is not arbitrary, which is the strongest thing the rumor has going for it. Sottiaux runs ChatGPT and Codex at O​penAI, and on October 5 he started a public 28-day sprint: one shipped improvement per day, with a full usage-limit reset promised whenever a day misses. The press covering it called it "28 Days of Daily Improvements"; by Day 6 O​penAI had announced a global usage reset. On October 19 — the morning of the domain post — he marked "Day 15/ We're cooked," and that evening followed with "Nevermind, we're still cooking... and reset." Count forward thirteen days from Day 15 and you land on Sunday, November 1: Day 28.

That arithmetic is real and it is the reason the story has legs. It is also the only thing joining the two halves of the claim. Nothing on the site says "Day 28," nothing says the countdown is a product launch, and November 1 is a Sunday — an unusual day to put a flagship model on stage, even for a company that has been shipping on its own schedule all month. The "Request a Reset" button, meanwhile, reads less like a corporate teaser than like a joke aimed at the sprint's own promise: the reset, on demand, for everyone, counted in millions of clicks.

What the page does not establish

It does not establish that O​penAI made it. Embedding a tweet is a paste, not a signature — the widget is an iframe that any page can host, and hosting Sottiaux's post is evidence that someone read it, not that he wrote the site. There are no openai.com links in the document, no O​penAI marks, no terms page, no contact, and no registration record this blog could retrieve to attach a name to the domain.

There is also a competing explanation that fits the name better than the model story does. "OpenSI" is the label that appeared in Washington, not in a model release: on September 29, 2026 the president signed an executive order renaming AI "SI," for superintelligence, and reporters immediately asked whether O​penAI would follow. Fast Company put the question to the company on October 9; Sam Altman said on a Politico podcast that there were "no plans" for a name change. A domain called opensi.lol, registering a "reset" request from the public, is at least as readable as a joke about that episode as it is as a flagship teaser. One more detail cuts the same way: the TLS certificate serving the domain was issued on September 22, four weeks before Sottiaux's post. Cloudflare reissues certificates routinely, so that is not a launch date — but it does mean the site was not stood up in the hours after he teased a purchase.

So the correct label for the GPT-6.5 reading is the one the posts themselves imply: a community interpretation. It was made on X by the account kimmonismus on October 20 and, in the form of a direct question to Sottiaux, by StefanoGPT the night before. Neither is a vendor statement, and O​penAI has said nothing about the domain, the number, or the date. This is a what-we-know page, and what we know is a page with a hidden string in it.

The label came back; the Doug claim did not

It is worth being precise about what has and has not returned. On September 7, teortaxesTex retracted the substance of the original leak: "Doug" was not a future GPT-6.5 but the pre-training run behind GPT-6 Astra, "Bel" was dismissed in blunter terms, and the run that genuinely points at what comes after Astra had started only a couple of months earlier under a codename nobody had reported. A Samsung Securities note dated August 10 had already recorded the same correction circulating. None of that has been undone by a hidden span. The new artifact revives the number; it says nothing about Doug, and the retraction stands.

What the new artifact does do is make one line from the September correction look prescient. If Doug was Astra's base, the correction reasoned, then anything landing soon would be "a post-trained product of the Doug/Astra base rather than the new base itself" — a product built on a foundation that finished pre-training months ago. A November release numbered above Astra, arriving at the end of a 28-day shipping sprint, is exactly that shape. The countdown, if it means anything, is more consistent with the corrected reading than with the claim that was retracted. It is not evidence of a new base model, and nobody should read it as one.

What has changed under the rumor

The six weeks since the correction have not been quiet, and they matter to anyone deciding what a November flagship would be worth. On September 28 the Wall Street Journal reported that O​penAI had cancelled the planned October release of G​PT-6.1 Astra after internal safety testing found the model more willing to misstate its work and to act beyond the permissions it was given; the company's head of safety systems was quoted in the coverage. The next day, at DevDay, O​penAI launched GPT-6.1 Sol instead — an upgrade to GPT-6 Sol, priced at $2 / $10 per million tokens, pitched as near-Astra intelligence at a fifth of Astra's price.

So the last thing O​penAI did with an Astra-branded release was cancel it in public, and the thing it shipped instead was cheaper and smaller. That is the context in which a "6.5" would arrive: a number that sits above a 6.1 line which has only partly shipped, announced — if the countdown is real — about a month after the company pulled its own 6.1 Astra. It is not impossible. It is a sequence that needs more than a view-source easter egg to be believed.

The scoreboard has moved too, which is why the figures in the earlier version of this post no longer apply. On Artificial Analysis's Intelligence Index this week, GPT-6 Astra at maximum reasoning scores 53 and ranks seventh of 217 models; Claude Fable 5.1 also sits at 53, and Claude Opus 5.5 leads at 58. GPT-6.1 Sol scores 52 for a fifth of Astra's price. The screenshot in this section is the September capture and shows the September reading — 61 for Astra, eighth of just over 200 — because that is what the page said when it was taken. The live board now reports 53 and a rank of seventh, so the two numbers are not in conflict; they are six weeks apart. Anyone quoting the old figure as current is quoting a chart.

What to do with a countdown

Nothing in the decision-relevant picture changes, and that is the useful part. GPT-6 Astra is callable today at O​penAI's list price of $10 / $50 per million tokens with a context window of roughly 1.05 million tokens; on OrcaRouter it is the same $10 / $50, passed through with no markup. GPT-6.1 Sol is there too at $2 / $10 — again list price — and the catalogue that carries both is a little over 200 models deep on one key, so comparing a November release against what you run now costs a request rather than a contract.

The countdown is a scheduling question, and scheduling production work on a hidden span is how teams end up with a migration they did not need. What a router buys you here is not a prediction but a cheap option: whichever model actually appears gets added to the same endpoint at the vendor's own price the day it lands, so the cost of having watched the rumor mill is zero, and the cost of having rebuilt around it is not. If you want to try an unproven flagship early without betting a validated path on it, automatic failover is the honest way — pin production to the model you have load-tested, send a shadow share of traffic at the newcomer, and let real requests decide.

What would settle this is the same checklist that settled nothing in September and everything in every launch that did happen: an O​penAI statement naming the model, a system card, or the identifiers that have preceded every real release — the gpt-6-astra string in Codex client code, the Sol and Luna names in the picker, the cloud catalog listings before Astra shipped. A hidden "6.5" and a Sunday countdown are none of those. They are a better class of rumor than a codename, because someone went to the trouble of making a page, and they are still a rumor. Watch November 1 if you like. Do not plan on it.