Redditor claims OpenAI silently downgrades older models, plans switch to DeepSeek and Kimi

Otheruser337 · reddit · 2026-09-30

The poster argues the new GPT-6 Sol/Luna models are mediocre 'sidegrades' versus the 5.6 generation — weaker at reasoning and planning despite better cost-efficiency — and coins a 'permaspike effect': strong models allegedly being intentionally toned down to save compute and make room for new model testing. Claiming OpenAI silently degrades older models to force migration, they plan to switch to Chinese models, using Kimi K2.7/K3 for coding and DeepSeek 4.1 Flash for planning and error-fixing. Unverified personal observation.

Original post →

More from Models

Models channel →