GPT-6.5 "Doug" Leak, Corrected: Doug Is GPT-6 Astra's Base Model

The shortest leak cycle this blog has tracked lasted about twelve hours. On September 7, the independent model analyst teortaxesTex predicted that OpenAI's next flagship would be branded GPT-6.5, that its internal codename was the reported pre-training run "Doug," and that it would be "something different… not just bigger" — a change in kind, not a scale-up. Within about twelve hours he had walked it back. After investigating, teortaxesTex says he had the wrong model attached to the wrong name: "Doug" is not a future GPT-6.5. In his corrected reading it is the base model of GPT-6 Astra — the flagship OpenAI shipped on September 3 at $10 / $50 per million tokens with a roughly 1.05-million-token context, whose standard tier reached general availability on September 5. The giant pre-training run the rumor mill spent August chasing was, on this account, the foundation under the thing that is already in production.

This page is a what-we-know tracker, so the labels have to move with the correction. GPT-6 Astra is real, independently measured, and callable; the codename story around it is not. teortaxesTex's follow-up is his own account of what he learned, not an OpenAI statement, and no artifact for "Doug" — no identifier, no model-picker entry, no cloud-catalog listing — has appeared under any interpretation. What the correction resolves is a misreading that the earlier version of this post repeated. Read what follows as a rumor trail correcting itself, with every claim still traceable to named or anonymous accounts rather than to OpenAI.

The claim, and the correction

teortaxesTex's original post, read strictly, had three parts: the next OpenAI model will carry the number GPT-6.5; its codename is Doug; and it represents a change in kind rather than a scale-up. The first two parts were naming hypotheses that lined up with the existing codename reporting. The third is the part that made the post travel, because it reached past "when" — the question the rest of the rumor mill was arguing about — and made a claim about "what." A successor that is different rather than bigger would not merely reset OpenAI's position on the independent leaderboards; it would change what the next generation of the frontier is.

His follow-up retracts most of that, and the correction has three parts of its own. First, "Doug" is the base model of GPT-6/Astra — not a separate, larger model waiting in the wings, but the pre-training run that the already-shipped Astra was built from. Second, "Bel," the reported 10-trillion-parameter successor, is — in his phrase — "insider bro bullshit." Third, the pre-training run that genuinely points at what comes after Astra started "maybe less than two months ago" under a different codename that nobody has reported. All three are teortaxesTex's statements about what he found, in the same register as the original leak: useful as a signal about where the rumor trail stands, worthless as a document. Nothing about them is an OpenAI confirmation, and he adds that none of this changes much in his view — the strategic point of the original post, he is saying, survives the name change. That is the part worth testing below.

Where the "Doug" reading came from — and how it got corrected

The future-model reading did not start with teortaxesTex, and it is worth being precise about its origins because the earlier version of this post leaned on it. On August 7, research firm SemiAnalysis published a memo — dated July 9 — telling institutional clients that OpenAI had "overcome their pre-training issues" and that "a much larger model code named 'Doug' is actively in the works." Two days later the X account ChrisGPT filled in the shape: Doug was not the same model as GPT-6 (which most accounts took to be Astra), it was OpenAI's largest pre-training project to date, and it could arrive as soon as November. On the strength of those two reports, "Doug" became the handle for a big, separate, year-end thing.

The reading that Doug is instead the base under Astra is not new either — and that is the strongest reason to treat teortaxesTex's correction as more than a whim. A Samsung Securities market note dated August 10, a month before his follow-up, already recorded the correction circulating in the community: some accounts read Doug as a separate model to appear after Astra, the note said, but it was "also corrected that Doug is the pre-training model that serves as the basis of Astra," with the caveat that the exact relationship remained unconfirmed. The Bel-era reporting was split on the same point: the anonymous account @synthwavedd's August 25 post described a finished run called "Bel" that some retellings made Doug's successor and others made "probably the base for GPT-6." The family tree was never consistent across accounts. teortaxesTex's follow-up simply picks the side the Samsung note had already recorded and drops the Bel layer entirely.

Read through the correction, the SemiAnalysis and ChrisGPT reports were tracking a run in its final stages rather than a future project. A base model that is "actively in the works" in early July, that clears an August safety review, and that ships as a product on September 3 is a base model whose pre-training finished months earlier — with the summer spent on reinforcement learning, alignment, and the rollout pause that followed Astra's "Critical" cybersecurity rating under OpenAI's own Preparedness Framework. That timeline fits Doug-as-Astra's-base. It does not fit Doug-as-a-November-launch. If teortaxesTex is right, the November window belonged to a misread run, and the reporting that promised a separate, larger Doug simply had the wrong object in view.

What survives the correction

The strategic thesis survives in a reframed form. SemiAnalysis's underlying argument — that OpenAI's gains since GPT-4o have come mostly from post-training, reinforcement learning, and inference-time compute on an older base, and that those returns diminish without a better foundation — is independent of which codename the foundation carried. If Doug is Astra's base, then the base-model reset the memo predicted has already happened and already shipped: GPT-6 Astra is the first product of restarted, corrected pre-training, not another polish of the GPT-4o-era brain.

That makes the measured record more interesting, not less. The reset's first product is not the world-beater the "biggest potato" reporting promised. On Artificial Analysis's Intelligence Index, GPT-6 Astra (max) scores 61 — about eighth of more than 200 models — while Claude Fable 5.1 tops the board at 66. If that is what a restarted base buys today, then "different, not just bigger" was never a description of what Astra already is; it was a hope about what the next base — the run that started less than two months ago under an unreported codename — will eventually produce. The energy of the original leak belongs to that future run, not to the codename everyone was arguing about.

Two further implications follow if the correction holds, and both are worth stating as inference rather than fact. First, the "successor is very soon" thread — Sam Altman's Axios comment that "much, much, much more capable models are coming soon," and the circulating clip in which he said the model OpenAI "recently talked about pausing" was "a future model" — cannot be about the fresh pre-training run; a run begun in mid-July is not a product in September. If a successor does land soon, it will be a post-trained product of the Doug/Astra base rather than the new base itself. Second, "Bel" was always the part of the story most exposed to exactly this kind of dismissal, which is the next section's subject.

Why "Bel" was the weakest link

The Bel report, from the anonymous account @synthwavedd on August 25, had the surface shape of a major leak: OpenAI had "finished" a giant pre-training run of more than 10 trillion total parameters, described as the successor to Doug and, on some readings, as the base for the GPT-6 generation. It never had the substance. It traced to a single anonymous post; no second source, no identifier, no picker entry, no cloud listing ever appeared; and the retellings contradicted one another on the most basic question of whether Bel came before or after Doug — a sign that the account had no fixed referent. The timeline was also incoherent on its face: a base model cannot "finish" pre-training at the end of August and also be the base of a model that had already been built, safety-reviewed, and readied for its September 3 launch.

teortaxesTex's dismissal — "insider bro bullshit" — is blunter than the reporting ever was, but it points at the same evidence. The blog's earlier successor explainer flagged the Bel/Doug inconsistency the week the report landed, and the Astra launch explainer carried the same caution. What the correction adds is a reason to stop hedging: if the real successor pre-training began less than two months ago under a codename no one has reported, then the "finished 10-trillion-parameter Bel" story was not merely unverified — it was describing something that, on the corrected timeline, could not have existed.

What would settle it

The checklist is unchanged, and the correction raises the bar rather than lowering it. Every OpenAI leak cycle this blog has tracked left a verifiable object behind before the announcement — the gpt-6-astra identifier in Codex client code, the Sol/Terra/Luna names in the GPT-5.6 model picker, the Azure and Bedrock catalog listings before Astra. Nothing in the Doug story has produced any of that, and the correction does not change it: even if Doug is Astra's base, the name is still an internal handle that no artifact has ever confirmed. What would settle the corrected reading is an OpenAI statement, a system card that names the base, or an insider account with a document — none of which exists. Until one does, the accurate label for "Doug is GPT-6 Astra's base model" is: the analyst who first called it GPT-6.5 now says so, a month-old research note recorded the same correction in the community, and OpenAI has said nothing.

The schedule question is where the correction bites hardest. If the genuinely-next pre-training run is only two months old and unnamed, then every date attached to "Doug" in the reporting — ChrisGPT's November window, the "very soon" successor expectations — was attached to the wrong run. What a reader can actually rely on is narrower: OpenAI has shipped Astra, executives have said more capable models are coming, and the next base-model reset is evidently early. None of that is a release date, and the honest position is that the successor clock starts whenever an identifier or an announcement appears — not when a leaker names a codename.

What to do while the story resolves

None of the decision-relevant facts change on a correction, because the models that exist are the ones you can call. GPT-6 Astra is live, benchmarked, and available through OpenAI's own API and ChatGPT plans — and on OrcaRouter at OpenAI's list price, $10 / $50, with the ~1.05M-token context passed through at provider pricing and zero markup. If you want a second opinion on the frontier right now, Claude Fable 5.1 tops the independent index, and GPT-5.6 Sol remains the proven cheaper OpenAI path for long-context work.

The useful preparation is the same as it was before the correction, because the uncertainty has not shrunk — it has moved. A pass-through router makes whichever successor actually appears callable through the key you already use, at the vendor's own list price, the same day it shows up in a provider catalog: no second contract, no rewrite, and the price change live the moment the vendor posts it. And because a brand-new base model is exactly the kind of unproven thing you do not want to bet a production path on before it has been load-tested, automatic failover is the honest way to try it: pin production to the model you have validated, point a shadow share of traffic at the newcomer, and let real requests decide. If the corrected reading is right and "Doug" was just the base under a model you can already call, nothing about that setup was wasted; if the unnamed successor ships, the cost of having followed the rumor mill is zero.

The fair summary of the correction is a tighter version of the original uncertainty. teortaxesTex, the analyst who put "GPT-6.5" on Doug, now says he had it wrong: Doug is the base model of the GPT-6 Astra that already shipped, "Bel" was noise, and the pre-training run that matters is less than two months old and unnamed — a reading a Samsung Securities note had already recorded a month earlier, and that no artifact or OpenAI statement has yet confirmed. What survives is the strategic point, now pointed at the right target: OpenAI has restarted base-model pre-training, its first product sits eighth on the independent index, and the run that decides what comes next has only just begun. The defensible posture is unchanged — keep calling the model you have validated, keep the option to route around the next release open, and treat an announcement, not a codename and not a correction, as the only schedule that matters.