Hallucination leaderboard 2026: best model still wrong 4.2%, 52 of 124 models net-negative
GaryMarcus · x · 2026-11-19
Gary Marcus cites the latest 2026 hallucination benchmarks to explain why most companies report no AI ROI: fixing errors costs time and money. Key numbers: best public model Gemini 4 Argon hallucinates 4.3% (Gemini 4 Carbon hits 4.2% but isn't public); Claude Fable 5.1, the most accurate model at 68.4%, still hallucinates 70% of the time; GPT-6 Astra sits at 49.8%; 52 of 124 models on the AA-Omniscience board give more wrong answers than right ones; legal AI sold as hallucination-free got 17–33% of queries wrong.
More from Models
- Heavy testing finds Astra 6.2 High slightly worse than 6.0 and 2x slower — IridiumEagle · 2026-11-19
- Heavy Astra 6.2 testing still can't exhaust the $100 plan's usage quota — prd_008 · 2026-11-19
- Gemini 4 Carbon Deciphers an Unread 1466 Sforza Renaissance Cipher Unaided — fofrAI · 2026-11-19
- Gemini's mysterious "Hydrogen" selector appears to route to Gemini 3.1 Pro — Neestor00 · 2026-11-19
- Fable 5.2 vs GPT-6.2 Astra: 108M vs 24M tokens to build the same PS1-style game — heypearlai · 2026-11-19
- Same game prompt: Astra ships for $17 vs Fable's $65 — 3x cheaper, fewer bugs — heypearlai · 2026-11-19