OpenAI's rumored 10-trillion-parameter GPT-6 Astra ships in August, five times GPT-4's size, as Anthropic and a weakened Google respond.
OpenAI's rumored 10-trillion-parameter GPT-6 Astra ships in August, five times GPT-4's size, as Anthropic and a weakened Google respond.

OpenAI's rumored 10-trillion-parameter GPT-6 Astra ships in August, five times GPT-4's size, as Anthropic and a weakened Google respond.
OpenAI is preparing to ship GPT-6, a rumored 10-trillion-parameter model five times the size of GPT-4, in August, ending a two-year pretraining drought as Anthropic and a Google losing its top researchers respond.
The release carries an unprecedented safety designation. OpenAI said in a company post that Astra is now the first system it treats as potentially "Critical" for cybersecurity under its Preparedness Framework, pausing internal activities that do not yet meet strengthened security controls.
AI insider reporter ChrisGPT, who first disclosed the timeline, said the model was built after OpenAI reallocated compute nine months ago, concerned that Anthropic had taken a lead in pretraining. The 10-trillion-parameter count would mark a fivefold jump from GPT-4's estimated 1.8 trillion, with a context window that could reach 1.5 million tokens.
The stakes extend beyond benchmarks. Alphabet shares fell more than 4 percent on Aug. 5 after chief scientist Jeff Dean and three core researchers left to found Discovery Loop and Nobel laureate Demis Hassabis stepped down as DeepMind CEO, thinning the bench of the only lab that matched OpenAI's scale.
A Two-Year Pretraining Drought Ends
Since GPT-4o shipped in May 2024, OpenAI has not completed a full-scale pretraining run of a next-generation frontier model. The o1, o3, GPT-5 and GPT-5.5 releases all refined the GPT-4o-era base with post-training, reinforcement learning and inference-time compute. The pretraining bottleneck forced projects like Orion to be downgraded after underperforming expectations.
The new run, codenamed Garlic, was a validation experiment after chief research officer Mark Chen told staff the company had solved key pretraining problems. Doug, the year-end model, scales that work further and is expected to train on Nvidia's Vera Rubin chips, according to SemiAnalysis and frontier-information analyst Pankaj Kumar. Doug is not GPT-6 — it is a separate, larger pretraining effort.
Anthropic's Counter-Launch and Google's Thinning Bench
Anthropic is not waiting. The company plans to release Fable 5.1 in August at the same price as Fable 5, a direct counter to Astra. The timing suggests a deliberate strategy: Opus 5 was held back at launch so Fable 5.1 could serve as a surprise intercept once GPT-6 appears.
Google, meanwhile, is losing the people who built its advantage. On Aug. 5, Jeff Dean — a 27-year veteran who co-founded Google Brain and helped design TensorFlow and TPUs — left with Sanjay Ghemawat, Oriol Vinyals and Quoc Le to found Discovery Loop, a public benefit corporation automating the scientific method. Alphabet is a founding investor and cloud partner. Hassabis, who won a Nobel Prize in 2024 for AlphaFold, handed daily operations to Koray Kavukcuoglu and became chairman and Alphabet's chief scientist.
The departures compound a summer of churn. Noam Shazeer, co-author of the 2017 Transformer paper, joined OpenAI in June, and Nobel laureate John Jumper moved to Anthropic — a wave that wiped $270 billion off Alphabet's market cap in June. Gemini 3.5 Pro, slated for a June launch, remains unreleased, with Bloomberg reporting coding performance fell short of internal targets.
The race is now effectively two-horse. Anthropic, valued at $965 billion, holds Fable 5; OpenAI counters with 5.5, 5.6 and the incoming 6. Alphabet's Q2 revenue rose 24 percent to $119.8 billion with Cloud up 82 percent, but the company turned cash flow negative for the first time on record after raising capex to $195-205 billion. Nvidia stands to gain from Doug's Vera Rubin demand, while Microsoft's in-house MAI-Cyber-1-Flash shows hyperscalers are building defense models rather than renting frontier systems. The question for investors is whether OpenAI's new engine — and the year-end Doug — can outrun Anthropic's intercept, and whether Google's remaining bench can ship Gemini 4 before the gap widens further.
This article is for informational purposes only and does not constitute investment advice.