AI news · leaked Aug 2026

OpenAI's Secret Model "Doug" Just Leaked

The Doug Leak Briefing

OpenAI's secret model just leaked.

It's codenamed Doug — and based on what's been reported, it's the biggest AI model OpenAI has ever built.

It's not even alone: a second model is already finished, and it could ship this month.

Behind both of them is a wild backstory — the most famous AI lab on Earth was stuck for two whole years.

Stick with me, because how they got unstuck changes what you should build next.

One leaked sentence "pre-training issues… overcome" Two frontier models Astra (finished) · Doug (in the works)
The sources — read them yourself ↓
01 · What happened

A respected research firm put it in writing.

On August 7, SemiAnalysis — the chips-and-labs research firm big investors pay for — published a newsletter quoting their own private client note from July 9.

"OpenAI has overcome their pre-training issues, and a much larger model codenamed Doug is actively in the works."

— SemiAnalysis, private client note (July 9, 2026), published in their newsletter Aug 7, 2026

Jul 9 Private client note: "Doug… in the works" Aug 7 The newsletter publishes the note Aug 8 Altman tweets about Astra Aug 9 ChrisGPT names "Doug" in public

Four dates, one story: a paid research note in July, published in August — then the CEO and the leakers filled in the rest within 48 hours.

02 · The leaker

Then a leaker gave it a name.

On August 9, the account ChrisGPT posted the codename in public — and the post pulled 1.5M+ views in a day.

Tweet 1 · the leak

"Doug" goes public

Chris says Doug is OpenAI's biggest pre-training run ever, as far as he knows — and claims it will make Claude's Fable 5 seem "primitive". Big claims, anonymous source: hold that thought for the skeptic's section below.

03 · What's solid, what's not

Two sources, two very different weights.

SemiAnalysis put its name on "Doug exists and pre-training is fixed" — the "biggest ever" and "primitive" lines come from an anonymous leaker, so treat them as claims, not facts.

NAME ON THE LINE Doug exists · pre-training fixed SemiAnalysis — paid research, reputation staked + the CEO's own Astra tweet CLAIMS, NOT FACTS "Biggest ever" · "makes Fable primitive" Anonymous leaker — unverified hold these loosely

Even if you bin everything the leaker said, "OpenAI has overcome their pre-training issues" is one of the most important sentences of the year — and there's a second model, already finished, that could ship this month. Keep scrolling.

04 · The backstory

Same brain since May 2024.

Pre-training builds the brain; post-training polishes it. Per SemiAnalysis, every OpenAI release since May 2024 has been the same brain, polished harder.

PRE-TRAINING — build the brain Enormous data on enormous compute POST-TRAINING — polish it Reasoning, instructions, behaviour ONE FOUNDATION — built May 2024 (GPT-4o) No true generational upgrade underneath since — per SemiAnalysis 4o o1 o3 GPT-5 5.5 5.6

o1 (Sep 2024), o3 (Apr 2025), GPT-5 (Aug 2025), 5.5, 5.6 — a two-year polish chain riding one foundation slab.

05 · Why they were stuck

The wall everyone believed in.

Pre-training kept breaking — and the industry concluded scaling had hit a ceiling.

GPT-4.5 lands soft — underwhelms everyone Orion reportedly downgraded for underperforming The bug reports of an obscure software bug stalling runs Two years no shippable full-scale pre-train completed INDUSTRY VERDICT: "pre-training hit a ceiling — all gains from post-training now"

Two years where the most famous lab in the world couldn't complete a shippable full-scale pre-train. Remember the verdict at the bottom — it turns out to be wrong.

06 · The trigger

Google punched, and OpenAI called Code Red.

Gemini 3 landed on November 18, 2025 — and by December 1, multiple outlets reported Sam Altman had declared "Code Red" inside OpenAI, pulling teams and moving resources.

Gemini 3 Nov 18, 2025 Google's haymaker CODE RED Dec 1 — Altman's alarm teams pulled, refocused Garlic Dec 2 — The Information the test run for the fixes Bugs fixed the "wall" problems reportedly SOLVED DOUG actively in the works A RIVAL'S PUNCH → THE HUNT FOR THE REAL PROBLEM → THE FIX → THE BIG SWING
07 · The fix

Garlic proved the problems were bugs.

One day after Code Red, The Information reported a new pre-trained model codenamed Garlic — with Chief Research Officer Mark Chen reportedly telling his team they'd solved the key problems breaking pre-training.

BEFORE THE FIXES Holding the knowledge took a much bigger, clumsier model AFTER THE FIXES (reported) Smaller models hold knowledge that used to need much bigger ones BURIED LINE: an even bigger model had already started, built on what Garlic taught. That model is Doug.

By January, SemiAnalysis wrote it plainly: pre-training issues resolved. By July 9: Doug actively in the works — the foundation scaling again for the first time in 2+ years.

"The wall wasn't a wall. It was a bug."
Before the leapfrogging starts

Don't weld your setup to one model.

Models are about to leapfrog fast — Doug reportedly by November, Fable 5.1 reportedly this month. December's top model won't be March's. The Agent OS inside the AI Profit Boardroom is built for exactly that: one dashboard, all your agents, model-swapping built in — the day Doug goes live, you swap it into your existing workflows in minutes.

The full Agent OS zip — one dashboard, all your agents, model-swapping built in
A 30-day roadmap — real lead-gen and time-saving agents on today's models
A video tutorial walking the whole setup, step by step
Daily updates — every new model wired in the week it ships
A room of 4,000+ founders running this exact stack
Get the Agent OS → Inside the AI Profit Boardroom · skool.com/ai-profit-lab
Set up in an afternoon · used in 38 countries · new models added the week they ship
08 · The first model

Astra is finished — and the CEO said so.

Astra is the first of the two models, reportedly done and sitting in safety review — and this part is confirmed by Sam Altman himself, not a leaker.

Tweet 2 · CEO-confirmed

Altman on Astra: powerful, for everyone, held for cyber safety

Sam says Astra is powerful, that he wants it generally available, and that its cyber capabilities mean they need a little longer to release it safely. Leakers expect it this month — many believe Astra is GPT-6, though that name is unconfirmed.

Tweet 3 · the poetic timing

Four years to the day since GPT-4

Greg Brockman marked that GPT-4 finished training exactly four years earlier — August 8. Four years: an estimated ~1.8T parameters then, a reported ~10T now.

The scale jump — if the reports hold GPT-4 (2022) 1.8T parameters — estimate Astra (2026) 10T parameters — REPORTED, unconfirmed 5×+ IF TRUE

~1.8T is the widely-cited estimate for GPT-4; ~10T for Astra is reported and unconfirmed. The chips say so on the chart, on purpose.

09 · The second model

Doug: the end-of-year swing.

Doug is the end-of-year model — Chris expects it no later than November, after government review and cybersecurity testing.

Tweet 4 · the compute angle

Reportedly training on chips you can't buy

Pankaj Kumar's roundup adds the hardware detail: Doug is most likely training on NVIDIA's next-gen Vera Rubin chips — hardware that isn't widely rolled out and that most companies can't buy yet. Reported, not confirmed.

ASTRA Finished · in safety review (CEO-confirmed) Leakers expect it this month ~10T REPORTED GPT-6? UNNAMED DOUG The end-of-year model · biggest pre-train (claimed) Then gov review + cyber testing NOV — REPORTED VERA RUBIN — REPORTED EVERY UNCONFIRMED NUMBER WEARS A "REPORTED" CHIP — ON PURPOSE
A 1.5M-token context — the whole shelf, held at once For a business, that's your entire operation — every email, client file and process doc — in mind at once JULY REPORTS — UNCONFIRMED

July reports say the next-gen foundation could reach a 1.5 million token context window — a full shelf of books held in mind at the same time.

10 · The skeptic's case

Now the honest case against all of it.

Before you get excited, here are the three best reasons to doubt everything above.

Skeptic point 1 · nothing is confirmed

OpenAI has confirmed nothing about Doug — no date, no price, no announcement, and nobody outside the lab has run a single prompt. The rumor pipeline is now doing pre-training too: 1.5M views for a model that doesn't officially exist.

Skeptic point 2 · big pre-trains have flopped before

GPT-4.5 landed soft, and Gemini 3.5 reportedly matched a months-earlier Anthropic model. Bigger has publicly failed to mean better — at both companies.

Skeptic point 3 · "makes Fable look primitive" ignores time

By November, Doug won't be competing with today's Fable 5 — Anthropic ships newer models by then, and Fable 5.1 is reportedly landing this month at the same price, timed to counter Astra. The target moves.

"Biggest ever" unverified claim "By November" leaker expectation "Makes Fable primitive" marketing-grade hype Hold claims loosely. Hold the DIRECTION firmly. THREE CREDIBLE THREADS — SEMIANALYSIS + THE INFORMATION + THE CEO — ALL POINT THE SAME WAY
10b · Three beliefs to drop

What the skeptic's case actually means for you.

Wrong: "Nothing about Doug is confirmed, so I can ignore this whole story."

Right: The specific claims are unverified — but the direction is backed by three credible threads: SemiAnalysis in writing, The Information's Garlic reporting, and the CEO's own tweet. You plan around directions, not press releases.

Wrong: "Big pre-trains keep disappointing, so Doug will flop too."

Right: Maybe it will — that's exactly why you own the workflows and keep the model swappable. Winners slot in, duds get skipped, and you profit either way.

Wrong: "I'll wait until Doug actually ships before I build anything."

Right: If it lands in November, that's exactly enough time to build systems on today's models — so you're upgrading on launch day while everyone else is starting from zero.

Don't take my word for it

Members post their wins every day — agency owners, ecom founders, course creators, solo operators across 38 countries. Real businesses, real numbers, in their own words.

Read the 158-page wins doc →
11 · The fourth signal

Google cracked — and the field narrowed to two.

On August 5, Jeff Dean left Google after 27 years to start a new lab, the same day Demis Hassabis stepped down as DeepMind CEO to become chairman — and Alphabet fell more than 4%.

Google · Aug 5 Jeff Dean out after 27 years → new lab "Discovery Loop" + 3 top researchers Hassabis: DeepMind CEO → chairman Alphabet −4% SemiAnalysis: "no longer a frontier lab" FIELD NARROWS OpenAI Astra (finished) Doug (in the works) BOTH ENGINES ON vs Anthropic Fable 5 (live) 5.1 (reported) COUNTER-PUNCHING TWO LABS, SWINGING AT EACH OTHER FOR THE REST OF 2026

SemiAnalysis called Dean the "undisputed GOAT of Google engineering". The lab that invented the Transformer is bleeding its best — leaving OpenAI and Anthropic to slug it out.

12 · What it means for you

Every release becomes a free upgrade — if you're set up right.

For two years one of OpenAI's two engines was off; if the reports hold, both now run at once — and the pace of releases speeds up for everyone.

YOUR WORKFLOWS — THE PART YOU OWN Client-email agent Lead-chasing agent Reporting agent MODEL SLOT — SWAPPABLE GPT-5.6 (today) Astra (reported this month) Doug (reported by Nov)
Old way Rebuild every launch
  • Your prompts and automations are welded to one model
  • A new model ships and you rebuild everything from scratch
  • You wait weeks before daring to touch the new thing
  • Every upgrade cycle costs you days of rework
  • The gap between you and AI-run businesses widens each release
New way Swap in minutes
  • Your workflows are the part you own — they don't change
  • The model sits in a swappable slot underneath them
  • The agent drafting client emails writes better the day the model ships
  • The agent chasing leads chases better — zero rebuild
  • Every release is a free upgrade to a business already running
12b · Already running

Members already run the workflows-you-own setup.

4,000+ Founders inside AIPB
400K YouTube subscribers
38 Countries · live members
163K X followers
29K Udemy students
A member's own Agent OS Mission Control running a Claude agent and a local Hermes agent side by side
Real member · built his own agent fleet — Claude + a local Hermes agent on one dashboard
Member win post: first automation with my brother — client invoicing that took 20 to 30 hours now entirely automated
Real member · first automation shipped — invoicing that took 20–30 hours, now entirely automated
Member win post: got the Obsidian vault wired onto his own Agent OS interface
Real member · wired the Obsidian memory vault onto his own Agent OS
Members post their wins in a 158-page doc — read it here →
13 · The clock

About three months — used right.

If Chris is right, Doug lands by November — exactly enough time to build real AI systems on today's models, so you're upgrading when it ships instead of starting.

Today → November (reported): ~3 months to build AUG — build on today's models SEP–OCT — systems running NOV — swap Doug in Upgrading beats starting. Every time.
Thinking it? "Doesn't running an Agent OS burn a fortune in tokens?"

That's the biggest myth about it. The everyday 90% runs on free local models on your own machine, free APIs slot in for more, and for frontier work it drives the CLIs you already pay for — your Claude subscription already includes the Claude Code CLI, and the Agent OS plugs straight into it, so you're never paying twice.

And inside the AI Profit Boardroom there are full token-optimisation tutorials, so usage drops to the bone and stays there.

Your three months start now

Be upgrading in November, not starting.

The AI Profit Boardroom is where you build the systems that make every model launch a free upgrade. When Astra ships, we walk through it live the same week. When Doug lands, it slots into workflows you already own.

The Agent OS with model-swapping — Astra and Doug slot in on day one
A 30-day roadmap — lead-gen and time-saving agents built on current models before November
4 weekly coaching calls going deep on releases the week they drop — Astra walkthrough when it ships, live setup questions answered
Daily tutorials — step-by-step videos as the tools change
The prompt library — cold outreach to onboarding, ready to run
A member map to find operators near you · someone online 24/7
A room of 4,000+ founders building through every launch
Get the Agent OS → Inside the AI Profit Boardroom · skool.com/ai-profit-lab
Set up in an afternoon · used in 38 countries · releases covered the week they drop
14 · Strip every rumour

Here's what's left.

The most famous lab in the world was stuck for two years.

The foundation wouldn't scale, and the industry decided the ceiling was real.

Then a rival punched hard enough to trigger Code Red.

They hunted the real problems, found bugs, and fixed them.

Now a respected research firm says their biggest model ever is actively in the works.

Doug by November — if the reports hold.

"The real question is what you'll have built by then."