2026 hallucination leaderboard: 141 of 202 models err more than they answer
gerardsans · x · 2026-10-25
Replying to "why does nobody talk about AI hallucinating anymore," this post links to suprmind.ai's continuously updated 2026 hallucination benchmark roundup. Key stats: among models answering ≥45% of questions correctly, Gemini 4 Argon has the lowest hallucination rate at 15.1% (not yet public); the best public model is Grok 4.7 at 29.3%. The most accurate model, Claude Fable 5.1 (67.2% accuracy), hallucinates 72.6% of the time; OpenAI's top GPT-6 Astra hits 51.3%. On the AA-Omniscience board, 141 of 202 models give more wrong than right answers. Legal/medical/finance AI sold as "hallucination-free" gets 17–33% of queries wrong.
Related event: AI Hallucinations Haven't Gone Away, Just Hidden by Agent Tooling(2 posts)→
More from Models
- AI companion app refuses to insult user even with explicit consent, sparking guardrail debate — Angaisb_ · 2026-10-25
- Casado: 'Abliterated models look behind the curtain' — a late-night vibe check — generativist · 2026-10-25
- GLM 5.3 Flash NVFP4 local-run repo assembled, tested up to 1M context (512K recommended) — HankYeomans · 2026-10-25
- Workday's returning CEO slashes AI agents in development from 50 to 20 — Scobleizer · 2026-10-25
- AI pulse: open models ~5 weeks behind frontier, Hermes agent merges 122 PRs in one day — Scobleizer · 2026-10-25
- Hugging Face researcher early impressions: Gemini 4 output "really nice" — eliebakouch · 2026-10-25