GPT-6 & Doug: What Actually Leaked (2026)

OpenAI’s next model leaked under the codename Doug — and the real story isn’t the hype, it’s one line from a respected research firm saying the pre-training wall that stopped OpenAI for two years is down. Here’s what’s sourced, what’s rumour, and what it means for you.

What actually leaked

On 7 August 2026, a research firm called SemiAnalysis published a newsletter. SemiAnalysis is one of the most respected firms tracking AI chips and AI labs — big investors pay them for research.

Inside it they quoted a private note sent to paying clients on 9 July, containing one line: OpenAI has overcome their pre-training issues and a much larger model codenamed Doug is actively in the works.

Two days later, on 9 August, an AI leaker who goes by Chris GPT on X posted about the same model — saying Doug will be OpenAI’s biggest pre-training run ever, and claiming it will make Claude Fable 5 seem “primitive”. That post did around 1.5 million views in a day.

But here’s the thing — throw out every word Chris said and the SemiAnalysis note alone is still a huge story.

Why “pre-training issues” is the important part

Every model has two stages. Pre-training is where it learns almost everything it knows by reading enormous amounts of data — building the brain. Post-training is where it’s taught to reason, follow instructions and behave — polishing the brain.

According to SemiAnalysis, OpenAI has been polishing the same brain since May 2024, when GPT-4o came out. Every big release since — o1 in September 2024, o3 in April 2025, GPT-5 in August 2025, then 5.5 and 5.6 — was built by polishing that same foundation harder. Better reasoning, more thinking time, smarter routing. But the brain underneath never got a true generational upgrade.

Why? Because pre-training kept breaking. GPT-4.5 came out and underwhelmed. The Orion project reportedly got downgraded after underperforming. Runs were plagued by bugs, including reports of an obscure bug in their training software that cost them months.

For two years the most famous AI lab on the planet couldn’t complete a full-scale pre-training run good enough to ship. The industry watched and concluded pre-training had hit a wall — that bigger doesn’t work any more, and all future gains would come from post-training.

Then Google forced the issue

1. 18 November 2025: Google released Gemini 3, strong enough to rattle the field.
2. 1 December 2025: multiple outlets reported Sam Altman declared a “code red” inside OpenAI. Teams pulled off projects, resources moved.
3. One day later: The Information reported OpenAI was working on a new pre-trained model codenamed Garlic.
4. Chief research officer Mark Chen reportedly told his team the company had solved what had been breaking their pre-training runs — actual bug fixes. The report said those fixes let smaller models hold knowledge that used to require much bigger ones.
5. January 2026: SemiAnalysis wrote plainly that OpenAI had resolved its pre-training issues.
6. 9 July 2026: the note confirming Doug is actively in the works.

Garlic was the test run. It proved the fixes worked. Buried in that same reporting was a line saying OpenAI had already started building something bigger based on what Garlic taught them — and that’s Doug.

So the wall that stopped OpenAI for two years wasn’t a law of physics. It was bugs. Fixable ones. And they’re fixed.

Related: the Garlic update and the Astra GPT-6 leak.

Astra comes first, and it’s confirmed

Doug is the second of two models in this story. The first is Astra, and this one isn’t a rumour.

Sam Altman tweeted about it himself on 8 August, saying Astra is a powerful model they want to make available to everyone, but because of cyber capabilities they need a bit longer to release it safely. So the CEO has publicly confirmed a finished model is sitting there, paused for security review.

• Leakers expect it this month, August 2026.
• Many believe Astra is GPT-6, though OpenAI hasn’t confirmed the name.
• Leaks put it at roughly 10 trillion parameters. GPT-4 was estimated around 1.8 trillion — more than five times bigger.

The timing is almost poetic: OpenAI president Greg Brockman just reposted that GPT-4 finished training on 9 August, exactly four years ago. Four years from 1.8 trillion to a reported 10 trillion.

What’s reported about Doug

• Chris expects it no later than November, because after pre-training finishes it reportedly needs government review and a pile of cybersecurity testing.
• Analyst Pangage Kuma pointed out Doug is most likely being trained on Nvidia’s next-generation Vera Rubin chips — hardware that hasn’t been widely rolled out yet.
• Reports from July said OpenAI’s next-generation foundation could reach a 1.5 million token context window.

1.5 million tokens is like handing someone a full shelf of books and having them remember every page while they answer. For a business, that’s your entire operation — every email, client file and process document — held in mind at once.

The sceptic’s case, which is fair

1. Nothing about Doug is official. No release date, no pricing, no announcement. Nobody outside OpenAI has run a single prompt through it. As one post put it, the rumour pipeline is now doing pre-training too — Doug has its own news cycle and doesn’t officially exist.
2. Big pre-trains have disappointed before. GPT-4.5 was hyped as a giant leap and landed soft. Gemini 3.5 reportedly came in around the level of something Anthropic shipped months earlier. A bigger brain doesn’t guarantee a better one, and it has failed publicly at both companies.
3. The “makes Fable look primitive” line doesn’t survive its own timeline. Chris pointed this out himself: by November, Doug won’t be competing with Fable 5. Anthropic will almost certainly have shipped something newer — Fable 5.1 is reportedly coming this month, priced the same as Fable 5, timed to counter Astra.

Hold the specific claims loosely. What you shouldn’t hold loosely is the direction: a named research firm says the pre-training wall is down, The Information’s Garlic reporting backs it from another angle, and OpenAI’s own CEO has confirmed a finished powerful model waiting on safety review. Three credible threads pointing the same way.

And a fourth signal: Google’s leadership cracked

On 5 August, Jeff Dean — Google’s chief scientist, co-founder of Google Brain and the man who started their chip programme — left after 27 years to start his own lab, Discovery Loop. He took three of Google’s top researchers with him the same day.

Demis Hassabis, a Nobel Prize winner, stepped down as CEO of DeepMind and moved to chairman. Alphabet stock fell more than 4% that day. SemiAnalysis went as far as saying DeepMind is no longer a frontier lab.

Sit with that. The lab that invented the transformer — the core technology inside every one of these models — is bleeding its best people. The field has narrowed to two: OpenAI holding Astra and Doug, Anthropic holding Fable 5 and reportedly 5.1.

What this actually means for you

If you run an agency, freelance, sell online or make content, it means the pace you’ve felt over the past year isn’t slowing down. It’s speeding up. For two years one of the two engines of AI progress was switched off at OpenAI and post-training carried everything. If these reports hold, both engines now run at once.

Here’s the practical read. Every release lands as a free upgrade for people who already have AI working in their business, and a bigger gap for people who don’t. If an agent already drafts your client emails, a smarter model writes them better the day it ships. If an agent already chases your leads, it chases them better.

The workflows are the part you own. The models are the part that keeps improving underneath you. Get the workflows right once and every release for the rest of this race — Astra, Fable 5.1, Doug, whatever follows — makes your business stronger without you lifting a finger.

That’s also why the timeline matters. If Doug arrives by November, that’s roughly three months — exactly enough time to get real systems built on today’s models so you’re upgrading rather than starting over.

FAQ

What is Doug?

A model reportedly in development at OpenAI, described by research firm SemiAnalysis as “much larger” and actively in the works. Nothing about it is officially confirmed by OpenAI.

Is Doug GPT-6?

Unclear. Many leakers believe Astra — the finished model awaiting safety review — is GPT-6, with Doug the bigger model behind it. OpenAI hasn’t confirmed either name.

How reliable is this leak?

It splits. The existence of Doug and the pre-training fix come from SemiAnalysis, a named firm writing to institutional clients. The superlatives come from an anonymous X account and are unverified.

What is Astra?

A reportedly finished OpenAI model. Sam Altman publicly confirmed on 8 August that it exists and that release is delayed for safety review over cyber capabilities. Leaks put it near 10 trillion parameters.

When would Doug arrive?

One leaker expects no later than November, because pre-training is reportedly followed by government review and cybersecurity testing. There is no official date.

Why does “pre-training is fixed” matter so much?

Because for two years the industry assumed pre-training had hit a ceiling. If it was bugs rather than physics, both engines of AI progress run at once again.

Should I wait for Doug before building anything?

No — that’s backwards. Build workflows on today’s models and every future release upgrades them automatically. Waiting just means starting from zero later.

What happened at Google DeepMind?

Jeff Dean left after 27 years to found Discovery Loop, taking three top researchers; Demis Hassabis moved from CEO to chairman; Alphabet stock fell over 4% that day.

The bottom line

Strip away every rumour and here’s what’s left: the most famous AI lab in the world was stuck for two years, a rival triggered a code red, they went hunting and found bugs rather than a ceiling, and a respected research firm now says their biggest model ever is actively in the works. Hold the GPT-6 Doug specifics loosely and the direction firmly — and build workflows now so whatever ships lands as an upgrade rather than a scramble.

About Julian Goldie

I run Goldie Agency, a 7-figure SEO agency, and teach this daily on a 400K+ subscriber YouTube channel. 240+ client projects on Upwork at a 100% job-success score, 10+ years through every major Google update. My systems are in the AI Profit Boardroom; my link building book is free here.