GPT 6 Astra cuts hallucination rate from 92% to 51%, claims biggest single-release gain
yanndubs · x · 2026-09-07
A Bridgemind AI post claims the new GPT 6 Astra dropped hallucination from 92% on GPT 5.6 Sol to 51% — billed as the biggest single-release improvement ever — while Anthropic's Fable 5.1 allegedly got smarter but lies more.
Quoting it, Yanndubs cautions that hallucination benchmarks must control for the number of claims a model makes, which neither plot does, so the numbers deserve scrutiny.
More from Models
- AI-generated engines and PCBs get fact-checked in viral AGI hype spat — teortaxesTex · 2026-09-07
- GPT-6 Astra scores 100% on music error detection, smashing prior 65% record — TheMoonMidas · 2026-09-07
- Astra rebuilds Penrose triangle and Ames room in Blender, showing real 3D reasoning — yshan2u · 2026-09-07
- Google ships Gemini 3.8 Flash Cyber, claims 2.6x more correct vulnerability patches — dl_weekly · 2026-09-07
- Debate grows over whether OpenAI and Anthropic will keep their best models in-house — steph_palazzolo · 2026-09-07
- Astra scores 279/280 with zero hallucinations on Signal65's PINNACLE enterprise benchmark — Tolopono · 2026-09-07