Title card for GPT-6.5 reading that the countdown was deleted three days before it ran.
Guides & Insights

GPT-6.5: The Countdown Was Deleted Three Days Before It Was Due

Author

Rowan Sterling

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

The page at opensi.lol was, until recently, the only concrete artifact behind the name GPT-6.5: a hidden element in the served HTML carrying a "6.5", a pair of coordinates on Fort Mason Center, and a time element counting down to 1 November 2026 at 10:00 AM Pacific. Fetching that page's source this morning returns none of it. No hidden aside, no "6.5" string, no countdown attribute, and no "requests worldwide" counter — the whole prior surface is gone, replaced by a single full-screen video that loops. Whichever way you read the last six weeks, the evidence did not get weaker. It was removed.

The opensi.lol page as served on 29 October 2026: a single full-screen looping video of a retro television on a wooden stand, with no counter, button or hidden countdown.

What is there now is a small, deliberate page. The body is one element: a video file, code>tv.mp4/code>, 7.3 MB, autoplaying and looping, with an audio-unmute-on-click handler and no other content. The document's title is just "OpenSI" — the September meta description, "Request a Reset," is gone with the button it described. The favicon links are cache-busted with code>v=20261019/code>, the date of Tibo Sottiaux's domain post. And the first line of the file, before the doctype, is a comment listing four handles: @gdb, @thsottiaux, @StefanoGPT, @poteto.

That comment is the only thing on the page that could be read as a claim about authorship, and it is not one. A comment is not a signature, and the four handles are the same set of accounts that were already being read as its audience in September. What has changed is that the page no longer argues its own case. It entertains.

What the artifact said, and what is left of it

The September version was checkable in a way that a video is not, and this blog checked it directly rather than taking a screenshot's word for it:

• A hidden aside with the id code>object-6-5/code>, invisible to a reader and visible in view-source.

• Inside it, a span reading "6.5"; a coordinate pair reading 37.8067, -122.4314, which resolves to Fort Mason Center on Marina Boulevard; and a time element whose datetime attribute was code>2026-11-01T10:00:00-08:00/code>.

• The -08:00 offset was correct rather than sloppy, because US daylight saving ends at 2 AM that same morning — 10 AM on 1 November really is Pacific Standard Time.

• A visible counter labelled "requests worldwide" that stood at 4,260,427 when this blog read the page, and a button reading "Request a Reset" that incremented it.

Every one of those facts was knowable by fetching a URL. Today, all but the coordinates are gone. The page still exists, still belongs to whoever owns the domain, and still says nothing about a model.

A six-row scoreboard for GPT-6.5 recording that the hidden string and countdown element were removed, that a 7.3 MB looping video replaced them, that four handles remain in the page source, and that no OpenAI statement exists.

Two readings, and no way to choose between them from outside

The removal is compatible with both of the stories that were already on the table, which is what makes it unhelpful as evidence and useful as a fact.

The first reading is that the countdown was never a launch timer. A teaser page that has served its purpose — passing around, being screenshotted, getting four million button presses — can be swapped for something more fun without the owner intending any signal. The video reads like that: it is a piece of content, not a countdown page. Under this reading, nothing about GPT-6.5 has changed except that the most-cited artifact for it stopped existing.

The second reading is that something moved. A page that had a hidden date in it and now does not is a page whose owner made a decision close to that date. That decision could be a pivot to a different announcement vehicle, a decision to delay, or a decision that the joke had run long enough. From outside, those three look identical, and none of them is a launch.

What you cannot do is read the deletion as confirmation of anything. A countdown removed is not a countdown satisfied, and the date is still three days away. If something does happen on Sunday, it will happen without the artifact that was supposed to announce it.

The arithmetic that made 1 November plausible still exists, and now points at nothing

The reason the date had legs was never the hidden element. It was the sprint. Sottiaux runs ChatGPT and Codex at OpenAI, and on 5 October started a public 28-day run: one shipped improvement per day, with a full usage-limit reset promised whenever a day misses. The press called it "28 Days of Daily Improvements." By Day 6 there was a global usage reset. On 19 October — 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 from Day 15 and you land on Sunday 1 November: Day 28.

That arithmetic is intact and still the strongest thing the rumour has. What has gone is the only page that pointed at the date. So the honest state of the claim is: a public sprint ends on a Sunday, a domain was bought on day fifteen, and a page that once counted to that Sunday no longer does.

There is a competing explanation that never depended on the artifact at all, and it fits the name better. "OpenSI" is the label that appeared in Washington, not in a model release: on 29 September 2026 the president signed an executive order renaming AI "SI," for superintelligence, and reporters immediately asked whether OpenAI would follow. Fast Company put the question to the company on 9 October; Sam Altman said on a Politico podcast that there were "no plans" for a name change. A domain called opensi.lol, serving a video and listing its audience in a comment, is at least as readable as a joke about that episode as it is as a flagship teaser. One detail cuts the same way: the TLS certificate serving the domain was issued on 22 September, four weeks before the post that was supposed to have prompted the purchase.

What the label's history does and does not contain

It is worth restating, because the deletion invites people to re-litigate it. In early September the analyst who first put GPT-6.5 on OpenAI's next flagship retracted the substance within about twelve hours: "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 10 August had already recorded the same correction circulating.

One line of that correction looks prescient rather than retracted. 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 November release numbered above Astra, arriving at the end of a 28-day sprint, is exactly that shape. It is not evidence of a new frontier base, and nobody should read it as one — no more than a hidden span was evidence of one.

The context a November release would land in

On 28 September the Wall Street Journal reported that OpenAI had cancelled the planned October release of GPT-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 next day, at DevDay, OpenAI launched GPT-6.1 Sol instead — an upgrade to GPT-6 Sol at $2.00 / $10.00 per million tokens, pitched as near-Astra intelligence at a fifth of Astra's price.

The last thing OpenAI did with an Astra-branded release was cancel it in public, and the thing it shipped instead was cheaper and smaller. A "6.5" would arrive into that sequence: a number above a 6.1 line that has only partly shipped, roughly a month after the company pulled its own 6.1 Astra.

Artificial Analysis leaderboard showing the leading models on its Intelligence Index, including Claude Opus 5.5 and GPT-6 Astra.

The scoreboard has moved too. On Artificial Analysis's Intelligence Index this week, GPT-6 Astra at maximum reasoning is listed at 57; Claude Opus 5.5 sits at 58, and Gemini 4 Carbon is on the board at 59. The 61 that was quoted for Astra in September is a chart from a different board state, not a rival number. Anyone comparing across those dates is comparing two snapshots.

The check that still works, with three days left

Every real release in this family has been preceded by identifiers rather than countdowns, and none of the checks below requires the deleted page:

• Client code. The gpt-6-astra string appeared in Codex client code before Astra shipped. A code>gpt-6-5/code> or code>gpt-6.5/code> string in a client bundle is the strongest pre-launch signal available, and it is checkable by anyone.

• Picker entries. Sol and Luna appeared as names in the product picker before any announcement. A picker entry is a distribution decision, which is later than a string in code.

• Cloud catalogue listings. Astra was listed in cloud catalogues before it shipped. A catalogue row with a price is where a rumour becomes purchasable.

• A system card or a vendor statement. Every real release has had one, and neither a hidden string nor a video is a substitute.

• The domain's behaviour. A page that changed once can change again. If code>tv.mp4/code> is replaced, if the comment's handle list grows, or if a countdown reappears, that is information.

What has already been answered, negatively, is the question the artifact was carrying: the page is not counting down any more. Anyone still refreshing it for a timer is refreshing for something that no longer exists on the page.

What is callable today

GPT-6 Astra is callable at OpenAI's list price of $10.00 / $50.00 per million tokens with a context window of roughly 1.05 million tokens; on OrcaRouter it is the same $10.00 / $50.00, passed through with no markup. GPT-6.1 Sol is there too at $2.00 / $10.00 — again list price — and the catalogue that carries both is 224 models deep on one key, so comparing whatever November brings against what you run now costs a request rather than a contract.

The scheduling decision is the one that actually costs money, and the deletion makes it easier rather than harder. Migrating production onto a hidden span was always the expensive version of hoping; now there is not even a span to point at. Waiting costs nothing, because whichever model appears gets added to the same endpoint at the vendor's own price the day it lands. 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.

Watch Sunday if you like. What the page can no longer tell you is that anything is coming — and the version where the date passes quietly is the one that leaves you nothing to unwind.

Compared in this article3

Detected from this article · Benchmarks: Artificial Analysis · updated daily