Leak claims OpenAI's next models: 5T-param sol on spud, 10T-param astra on doug pre-train
GregCook2011 · x · 2026-10-12
An unverified leak lays out OpenAI's alleged pre-training roadmap:
- spud was reportedly OpenAI's answer to being weak at pre-training but strong at post-training; it underpins sol, a 5 trillion-parameter model, alongside terra and luna in the same family.
- The next model, astra, uses an even larger pre-train codenamed doug, at 10 trillion parameters — reportedly the model OpenAI researchers are considerably more excited about.
- The author claims researchers see 15T as the magic number for recursive self-improvement, with both OpenAI and Anthropic having hit it, and reads Karpathy's recent posting activity as corroboration.
- Highly speculative in tone (IPO talk, hard takeoff); treat as rumor.
More from Models
- Claude Code Users Report Fast Mode Locked Despite Plenty of Remaining Credits — RileyRalmuto · 2026-10-12
- OpenAI's new Decisions API fails a simple coin-flip calibration test — JnBrymn · 2026-10-12
- Unverified leak claims Anthropic's next Claude is being tested in Claude Code, billed as SOTA — Eric520CC · 2026-10-12
- Prediction: Anthropic keeps public model lead until OpenAI's December release — imjustnewatai · 2026-10-12
- Ant Group Releases PolyOCR, Unified OCR Foundation Models with SOTA Results — GuangJian Team · 2026-10-12
- Dev vents: OpenAI models still suck at writing, why is that? — beechinour · 2026-10-12