Benchmarks are moving targets in 2026, and error rates shift wildly depending...

https://zaneznae304.lucialpiazzale.com/gemini-3-pro-hallucinated-88-on-aa-omniscience-is-it-still-usable

Benchmarks are moving targets in 2026, and error rates shift wildly depending on the test. Take HalluHard, which still clocks 30.2% failure rates even with live web access. If you are building for production, stop relying on generic scorecards