On 7 August 2026, one of AI's most credible research firms quietly confirmed that GPT 6 Doug is real — a much larger OpenAI model, actively in the works — and within two days an anonymous leaker had turned that one line into the loudest AI rumour of the year, so this page does one job: separate the verified, the credibly reported and the pure speculation. Nobody outside OpenAI has touched it; nothing here is my testing — just leak coverage, sourced claim by claim.
What Leaked, and Who Actually Said It
Two sources, completely different weight classes.
The serious one first. On 7 August, SemiAnalysis — one of the most respected research firms tracking AI chips and AI labs, the kind big investors pay for — published a newsletter quoting a private note they had sent paying clients on 9 July. The line: OpenAI has overcome its pre-training issues, and a much larger model codenamed Doug is actively in the works.
Then the loud one. On 9 August, an anonymous AI leaker posting under the name Chris GPT claimed Doug will be OpenAI's biggest pre-training run ever, and that it will make Claude Fable 5 seem "primitive". Roughly 1.5 million views in about a day. That is an anonymous account, so treat every word as a claim, not a fact.
The existence of Doug rests on SemiAnalysis; the hype rests on a leaker. Keep that split in mind.
The Two-Year Problem That Makes Doug a Big Deal
Here is the background SemiAnalysis lays out. Every model has two stages. Pre-training is where the model learns almost everything, reading enormous amounts of data on enormous amounts of compute — building the brain. Post-training is where you teach the finished brain to reason, follow instructions and behave — polishing the brain.
The uncomfortable part: OpenAI has been polishing the same brain since May 2024. GPT-4o was the last true foundation. o1 in September 2024, o3 in April 2025, GPT-5 in August 2025, then 5.5 and 5.6 — per SemiAnalysis, all built by polishing that same foundation. The brain underneath never got a true generational upgrade.
Because pre-training kept breaking. GPT-4.5 underwhelmed. The big Orion project reportedly got downgraded for underperforming. Reports described an obscure bug in OpenAI's training software stalling things for months. For two years, the most famous AI lab on the planet could not complete a full-scale pre-training run good enough to ship as its next frontier model — and the industry largely concluded pre-training had hit a wall.
If you want a business that upgrades the day a new model ships instead of being welded to the old one, the AI Profit Boardroom ships the Agent OS where swapping in a new model takes minutes, not a rebuild. → Get model-proof before Doug lands
Code Red, Garlic and the Bug Fixes
The turn starts with Google. On 18 November 2025, Google released Gemini 3 — strong enough that on 1 December, multiple outlets reported Sam Altman had declared a "code red" inside OpenAI.
One day later, The Information reported a new pre-trained model codenamed Garlic — and the key detail: OpenAI's chief research officer Mark Chen reportedly told his team the company had solved the problems breaking their pre-training runs. Actual bug fixes — and the report said those fixes let smaller models hold knowledge that used to require much bigger ones. Garlic was the test run proving the fixes worked. Buried in the same report: OpenAI had already started building an even bigger model based on what Garlic taught them. That model is Doug.
By January 2026, SemiAnalysis was writing plainly that OpenAI had resolved its pre-training issues, and their 9 July note confirmed Doug actively in the works. If the reporting holds, the wall was never a law of physics. It was bugs — fixable, and now reportedly fixed.
📺 Watch: GPT-6 + Doug: OpenAI's Secret AI
Astra: The Model Altman Has Already Confirmed
Before Doug there is a closer model, and its existence is not rumour. Astra is reportedly finished, and Sam Altman himself tweeted on 8 August that Astra is a powerful model they want to make available to everyone, but because of its cyber capabilities they need a bit longer to release it safely. So: CEO-confirmed publicly, trained, paused for security review.
Leakers expect Astra this month, and many believe Astra is GPT-6 — though OpenAI has not confirmed the name. Leaks put it at roughly 10 trillion parameters against an estimated 1.8 trillion for GPT-4 — more than five times bigger, though those are unverified leak numbers. Poetic timing: OpenAI president Greg Brockman just reposted that GPT-4 finished training on 9 August, exactly four years earlier.
Doug: The Reported End-of-Year Model
Doug is the end-of-year swing. The leaker expects it no later than November, because after pre-training it reportedly needs government review and cybersecurity testing before anyone outside OpenAI touches it. One tech analyst pointed out that Doug is most likely being trained on Nvidia's next-generation Vera Rubin chips — hardware not even widely rolled out yet; if accurate, chips most companies on Earth cannot buy.
July reports also said OpenAI's next-generation foundation could reach a 1.5-million-token context window — enough to hold an entire operation in mind at once: emails, client files, process docs.
📺 Watch: Qwen 3.8-Max, Fable 5 and GPT-5.6 Sol: Who Wins?
GPT 6 Doug: Every Claim, Sourced and Rated
The table I wish every leak thread shipped with. Confirmed: someone with their name on it said it publicly. Reported: a credible outlet or firm published it without official confirmation. Unverified rumour: anonymous accounts and leak numbers.
| Claim | Source | Status |
|---|---|---|
| Astra exists, is powerful, and is held back for cyber-safety work | Sam Altman's own post, 8 August | Confirmed |
| OpenAI overcame its pre-training issues; a much larger model codenamed Doug is in the works | SemiAnalysis client note, 9 July, published 7 August | Reported |
| Pre-training bugs solved; Garlic proved the fixes; a bigger model already underway | The Information, December 2025 | Reported |
| Doug likely trained on Nvidia's next-generation Vera Rubin chips | One tech analyst | Reported |
| Next-generation foundation could reach a 1.5-million-token context window | July reports | Reported |
| Doug is OpenAI's biggest pre-training run ever and makes Fable 5 seem "primitive" | Anonymous leaker Chris GPT | Unverified rumour |
| Astra is GPT-6 at roughly 10 trillion parameters, shipping this month; Doug by November | Leakers | Unverified rumour |
The Sceptics' Case
Three things stop me swallowing this whole.
- Nothing about Doug is officially confirmed. No release date, no pricing, no announcement — and nobody outside OpenAI has run a single prompt through it. One post summed it up: the rumour pipeline is now doing pre-training too. Doug has 1.5 million views and its own news cycle, and it does not officially exist.
- Big pre-training runs have disappointed before, publicly. GPT-4.5 was hyped and landed soft. Google's Gemini 3.5 reportedly came in around the level of a model Anthropic had shipped months earlier. Bigger brains do not guarantee better ones.
- The "makes Fable 5 look primitive" line fails its own timeline. By November, Anthropic will almost certainly have shipped something newer — Fable 5.1 is reportedly coming this month, priced the same as Fable 5 and timed to counter Astra. The target moves.
My position: hold the specific claims loosely, hold the direction less loosely. Three separate credible threads — the SemiAnalysis note, The Information's Garlic reporting and Altman's own Astra post — all point the same way.
📺 Watch: Kimi K3 VS Fable 5 VS GPT 5.6: Who Wins?
The Fourth Signal: Google DeepMind's Brutal Week
On 5 August, Jeff Dean — Google's chief scientist, whom SemiAnalysis calls the undisputed GOAT of Google engineering, co-founder of Google Brain who started their chip programme — left after 27 years to start his own lab, Discovery Loop, taking three of Google's top researchers the same day. Demis Hassabis, a Nobel Prize winner, stepped down as DeepMind CEO and moved to chairman. Alphabet stock fell more than 4% that day, and SemiAnalysis went as far as saying DeepMind is no longer a frontier lab.
Agree or not, the field is narrowing: OpenAI holding Astra and Doug, Anthropic holding Fable 5 and reportedly 5.1, swinging at each other for the rest of 2026.
What This Means for Your Business
So what does GPT 6 Doug mean if you run a business, not a research lab? Models are about to leapfrog fast — whatever is on top in December probably will not be in March. Every release lands as a free upgrade for businesses already running AI workflows, and a bigger gap for those that are not. If an agent already drafts your client emails, a smarter model writes them better the day it ships.
The workflows are the part you own; the models improve underneath you. That is the philosophy behind my Agent OS and the wider agentic operating system approach: build the system once, swap engines whenever a better one appears. It is why I keep a list of the best open-source models for Hermes Agent, and why Prime Agent was never welded to one provider — the GLM 5.5 lesson all over again.
A reported November Doug gives you roughly three months — exactly enough time to build real systems on today's models, so when the new ones land you are upgrading, not starting over. And whichever of Astra, Doug or Fable 5.1 ships first goes straight into my Goldie Bench side-by-sides, because leak numbers never survive contact with real tasks unchecked.
FAQ: Doug and Astra
Is GPT 6 Doug actually real?
Credibly reported, yes: SemiAnalysis told paying clients Doug is actively in the works. But OpenAI has announced nothing, so real means reported by a serious firm, not confirmed.
Is Astra GPT-6?
Many leakers believe so; OpenAI has not confirmed the name. What is confirmed, via Sam Altman's own post, is that Astra exists, is powerful, and is held for cyber-safety work.
When does Doug come out?
The leaker expects no later than November, after reported government review and cybersecurity testing. Treat that as rumour, not a date; Astra is nearer, expected this month by leakers.
Will Doug beat Claude Fable 5?
The "primitive" claim comes from an anonymous account, and by November Anthropic will likely have shipped Fable 5.1, reportedly landing this month at the same price as Fable 5. Wait for real side-by-sides.
What should I do right now?
Build workflows on today's models and keep them model-agnostic. Three months is enough to get real systems live, so every release becomes a free upgrade rather than a rebuild.
My Verdict
Here is where I land on GPT 6 Doug. Confirmed: Astra exists and is coming, on Altman's own word. Credibly reported: OpenAI fixed its pre-training problems and is building something much bigger. Unverified: every parameter count, every date, every "primitive". The two-year foundation drought looks like it is ending, Anthropic is loading its counterpunch, and the smartest move costs nothing — build the workflows now, and let the giants fight over whose engine runs underneath them.
If you want to be ready the day Astra or Doug goes live, check out the AI Profit Boardroom — inside you get the Agent OS with model swapping built in so Astra or Doug slot straight into your workflows the day they are available, a 30-day roadmap, four coaching calls a week, daily tutorials, and a community preparing for this exact wave. → Build your model-proof system today
