TL;DR: Multiple leakers, including AI insider ChrisGPT, report that OpenAI's paused model Astra is actually GPT-6 — with an estimated ~10 trillion parameters (more than 5× GPT-4's ~1.8T) — and that it will ship in August 2026 despite government review pressure. A second, much larger model codenamed "Doug" is expected by year-end, possibly trained on NVIDIA's next-gen Vera Rubin chips. Anthropic is reportedly readying Fable 5.1 as a direct counter. All of this is leaked and unconfirmed — but the strategic picture it paints lines up with everything else we know.
Table of Contents
1. The Leak: Astra Is GPT-6, Launching in August
On the morning of August 10, AI insider ChrisGPT posted on X that OpenAI will launch GPT-6 (codenamed Astra) in August — "even under the pressure of government review." Astra is the same model OpenAI urgently paused days earlier, according to the leak, and it's being positioned as the company's answer to the growing perception that it has fallen behind in pre-training.
The headline number: ~10 trillion parameters. For context, GPT-4 is estimated at roughly 1.8 trillion — so Astra would be more than 5× larger. OpenAI reportedly began reallocating compute toward this push about nine months ago, after realizing Anthropic had taken a lead in pre-training.
⚠️ Rumor watch: All of this comes from leaks — ChrisGPT, SemiAnalysis, and other unnamed sources — not from OpenAI. Treat the specific numbers as informed speculation. What's more credible is the strategic direction: OpenAI has been signaling for months that a major pre-training reset was coming.
This isn't the first time "Astra" has shown up. Earlier in August, reports surfaced that OpenAI's next-gen model Astra solved 10 long-standing open problems in mathematics — we covered that in depth in OpenAI Astra Solves 10 Unsolved Math Problems. If those proofs were generated with the same ~10T-parameter model, the leak is consistent: this is a genuine next-generation model, not another incremental post-training refresh.
2. Doug: The Year-End "Epic Beast"
The second leak is arguably the bigger story. ChrisGPT also named a model codenamed Doug — described as OpenAI's "most massive pre-training model to date," the kind of scale that reportedly makes Fable "look like a primitive artifact."
- Timeline: Needs pre-training plus extensive cybersecurity testing; possibly as late as November — i.e., "by year-end."
- Hardware: Speculative analysis (including by SemiAnalysis and Pankaj Kumar) suggests Doug is likely trained on NVIDIA's next-gen Vera Rubin chips, which haven't fully rolled out yet.
- Strategy: The playbook reads as: use the 10T GPT-6 to grab attention and users in August, then drop "Doug" at year-end to lock in the #1 position.
Intriguingly, the name Doug isn't new — SemiAnalysis mentioned it earlier, and leaks from July 2026 pointed to OpenAI's next pre-training base moving to ~10T tokens (up from the ~4T "Spud" base) with context possibly reaching 1.5M tokens, aimed squarely at Anthropic's Mythos/Fable line.
3. Why OpenAI Has to Ship Now: Two Years Without a New Engine
Here's the uncomfortable fact the leaks point to: since GPT-4o (May 2024), OpenAI hasn't completed a full-scale pre-training run of a next-generation frontier model. The o1, o3, GPT-5, and GPT-5.5 releases were all built on the GPT-4o-era base — post-training, RL, and inference scaling on top of an engine that hasn't fundamentally changed in over two years.
That's why projects like Orion were reportedly downgraded when they underperformed expectations, and why the pressure mounted through late 2025. Two catalysts accelerated the reset:
- Gemini 3's launch (late 2025) convinced Sam Altman the engine had to change — a new pre-training base codenamed Garlic emerged soon after.
- Chief Research Officer Mark Chen reportedly told the team the key pre-training problems were solved — Garlic was the validation experiment, and Doug is the real product scaled up.
If OpenAI really cracked the performance-degradation bottleneck and stabilized 10,000+ GPU (possibly 100,000+) clusters, then Doug represents something the field hasn't seen in a while: pre-training scaling delivering again, not just inference-time scaling.
4. Anthropic's Counterattack: Fable 5.1
Anthropic isn't waiting. Reports say the company is preparing to release Fable 5.1 in August, priced the same as Fable 5, positioned directly against Astra.
The read from industry watchers: this is a deliberate "racing strategy" (田忌赛马) move. Opus 5 was held back on purpose — Fable 5.1 is the reserve horse, ready to counter-attack the moment GPT-6 debuts. If Fable 5.1 lands with comparable benchmarks at launch-day pricing, Anthropic could blunt OpenAI's August moment.
Meanwhile, Google's talent exodus adds another layer: Jeff Dean — 27 years at Google — reportedly left with three core researchers to found a startup, and Nobel laureate Demis Hassabis stepped down as DeepMind CEO (moving to chairman) the same day. Alphabet's stock dropped over 4%. The lab that invented the Transformer is now exporting its top researchers to competitors.
5. What This Means for the Frontier AI Race
Putting the leaks together, the competitive landscape for H2 2026 looks like this:
| Player | Near-term | Year-end |
|---|---|---|
| OpenAI | GPT-6 / Astra (~10T params, August) | Doug (largest pre-training run, Vera Rubin?) |
| Anthropic | Fable 5.1 (August counter) | Mythos/Fable line evolution |
| Google DeepMind | Gemini 3 era; leadership exodus | Rebuilding after Dean/Hassabis departures |
Three takeaways for developers and businesses:
- Expect a pricing/quality shakeup in August. Two frontier releases in the same month means better models at competitive prices. If you're locked into one provider's API, this is a good moment to re-benchmark.
- Long-context is the battleground. The 1.5M-token context rumor (if real) would make agents dramatically more useful — worth planning your RAG vs. long-context strategy now.
- Pre-training scaling isn't dead. SemiAnalysis has argued all along that Scaling Laws still hold. Doug, if real, is the proof point: the "inference-time scaling" era was a detour, not the destination.
For the AI tooling ecosystem, this also matters: every agent framework we've reviewed — from Prime Agent to Agent Skills — runs on top of these frontier models. A genuine leap in base-model capability resets what agents can do, which is why this leak matters beyond the benchmark charts.
6. FAQ
Q: Is GPT-6 real, or just a rumor?
The August launch of a model called Astra/GPT-6 comes from leaks (ChrisGPT, SemiAnalysis, unnamed sources) and is not confirmed by OpenAI. The scale figure (~10T parameters) is an estimate. The strategic pattern — a long-delayed pre-training reset, a paused release, an Anthropic counter — is consistent with months of reporting, but treat specifics as speculation until OpenAI announces.
Q: What is "Doug"?
A reportedly much larger OpenAI pre-training model expected by year-end (possibly November), said to be the real product of the "Garlic" validation run and possibly trained on NVIDIA Vera Rubin hardware. SemiAnalysis had previously referenced the name.
Q: How big is 10 trillion parameters vs GPT-4?
GPT-4 is estimated at ~1.8T parameters, so ~10T would be more than 5× larger — a step change in the same league as GPT-3 → GPT-4, not an incremental bump.
Q: Should I switch AI providers because of this?
Not yet — it's unconfirmed. But August is a sensible time to re-run your own benchmark suite across OpenAI, Anthropic, and Google APIs, especially if you use long-context or agentic workloads. Prices often reset around frontier launches.