Benchmarks in 2026 are all over the place, and hallucination rates vary wildly...

https://www.pawn-bookmarks.win/in-2026-accuracy-is-just-marketing-noise-hallucination-rates-shift-wildly

Benchmarks in 2026 are all over the place, and hallucination rates vary wildly by test. Even with web search, the HalluHard benchmark shows a 30.2% error rate