gpt-6IMAGE
HalluHard hallucination benchmark: GPT-6 Astra leads all models
The author says GPT-6 Astra clearly outperforms other models, including Fable 5, on the HalluHard hallucination benchmark, both with and without web search.
Maksym Andriushchenko
@maksym_andrArchived 2026.09.18Originally posted 2026.09.18
Maksym Andriushchenko @maksym_andrOriginally posted 2026.09.18
💥 New results on HalluHard: GPT-6-Astra is significantly better than any other model, including Fable 5. This holds both with & without web search! Not many people are discussing hallucinations now, but they're still an important indicator of the lack of uncertainty awareness.
x.com/maksym_andr/status/2100641789363306643Archived 2026.09.18 10:08