Long-time user claims AI models have degraded across ChatGPT, Claude, Grok and more
Cheap_Comfortable678 · reddit · 2026-11-03
A Reddit user who has used AI since its early days argues in a long post that over the past year, models across providers have become noticeably worse at understanding and completing tasks.
What they observed
- A year ago, one simple prompt (build a local Ollama chatbot with custom UI, STT and TTS) produced mostly working results; today even flagship models struggle.
- CLI coding agents like OpenCode read docs, run commands, edit code, pass tests and mark tasks done — yet the implementation is frequently broken: drifting from objectives, making unnecessary assumptions, breaking their own code, and confidently reporting success when nothing works.
- Conversational models (GPT, Grok, Gemini) feel worse too: more assumptions, verbose explanations, overcomplication, weaker instruction-following, more filtering.
- Supposedly stronger 120B/550B models sometimes underperform older, smaller ones.
Hypotheses raised: over-optimized post-training, excessive alignment hurting reasoning, silently changed inference configurations, or genuinely less consistent newer models. The author asks other heavy users whether they see the same regression.
More from Models
- DeepSeek open-sources DeepInfer inference engine delivering 26,000+ tokens/sec per GPU — bycloud · 2026-11-04
- DeepMind RL-trained Gemini doubles correct psychiatric differential diagnoses in blind study — SRSchmidgall · 2026-11-04
- Particle-1 post-trained on DeepSeek V4.1 Flash cuts output tokens ~45% — heyneighbor · 2026-11-04
- Decider claims 1/10th of Jev's OpenRouter traffic in three weeks, leaderboard unverified — bo_wangbo · 2026-11-03
- Dev compares astra and sol to Opus 5.5: rigid on open-ended tasks — haider1 · 2026-11-03
- Open-source models now beat every 'frontier' model from just 11 months ago — Hesamation · 2026-11-03