on average it is fine. On difficult tasks it still does hallucinate. - like decision based on huge policy documents ( context reaching around 900k)

A lot less these days tho.