Benchmarks are all over the map in 2026. HalluHard shows a 30.2% error rate...

https://alexisqoem867.trexgame.net/the-illusion-of-provenance-dissecting-the-citation-failures-of-perplexity-and-copilot

Benchmarks are all over the map in 2026. HalluHard shows a 30.2% error rate even with web search enabled. You cannot just pick a single score and trust it for your stack