One in Five AI Benchmarks Regresses in Newer Models
An analysis by Paradigm 3 of over 5,000 model-benchmark pairs found that about one in five benchmark scores decline across successive model versions, with median drops mostly under 8%, showing frontier AI models can sometimes get worse.
2026-08-27 ~ 2026-08-27 · 2 related posts
- One Fifth of AI Benchmarks Get Worse in Successive Models — gleech · 2026-08-27
- Study Finds 1 in 5 AI Benchmarks Regress in Newer Models — gleech · 2026-08-27