Rumor math: OpenAI's Astra may be just 2-3x Kimi K3, gains from RL not scale
teortaxesTex · x · 2026-09-07
Twitter user teortaxesTex offers an unverified back-of-envelope analysis of OpenAI's rumored Astra model:
- Astra trails Fable badly on knowledge-heavy evals and is only marginally better than Opus or Sol; it's a mediocre creative writer, and total scale correlates strongly with writing ability
- Serial depth suggests ≤2x GPT-4 (≤240 layers); looping rumors would mean unique parameters are below per-forward-pass count
- Conclusion: Astra is plausibly a 2-3x larger Kimi K3 with modest algorithmic progress, with most of its edge from OpenAI's reasoning data plus heavy RL — including extreme RL scenarios that now cash out as ImageGen + Blender workflows
The author estimates >50% of compute went to RL and >150T training tokens. All figures are speculation, not confirmed.
Related event: GPT-6 Astra Rumors Swirl Over 1.2T Active Parameters(2 posts)→
More from Models
- GPT-6 Astra fails badly at dialog-centric interpersonal game's first-time user experience — msew · 2026-09-07
- GPT 6 Astra Strains Under Heavy Usage as Users Pushed to Switch to 5.6 sol — vista8 · 2026-09-07
- Rumor: Gemini 3.5 Live may be coming soon, possibly Pro-subscription only — Able-Line2683 · 2026-09-07
- ACX: Claude Fable 5.1 Cracks 370-Year-Old Cipher, Apollo Research Expands Hiring — Astral Codex Ten · 2026-09-07
- OpenAI's Astra Decompiles BIOS to Nail EC Backlight Control Other Models Couldn't — anom604 · 2026-09-07
- User reports Astra high feels better and burns far less weekly quota than Sol — sweetbeard · 2026-09-07