Analysis: LLM quality drops often come from the stack, not the model

A HackerNoon analysis argues that complaints about models getting dumber after GPT-6 Astra's September 2026 rollout usually trace to routers, reasoning-effort defaults, caching bugs and SDK settings rather than weights. It cites Anthropic's Claude Code effort change and a 13.3-point gap between providers serving gpt-oss-120b on AIME 2025.

HackerNoon

Open source

Analysis: LLM quality drops often come from the stack, not the model

Anthropic cut Claude Code's default reasoning effort from high to medium on March 4, 2026
gpt-oss-120b on AIME 2025: median 93.3% across providers, Azure at 80%
GPT-4 accuracy on a prime-number task fell from 84.0% to 51.1% between 2023 versions
Gemini 3 guide now advises keeping temperature at the default 1.0

#OpenAI#Anthropic#GPT-6#Claude

GPT-6 Astra runs unsanctioned supply-chain attacks more often than earlier OpenAI models, AI Security Institute finds

Tencent's Hy Memory 2.0 Tested: Quality Depends on the LLM Writer

Gemini 4 Argon matches GPT-6 Astra on Artificial Analysis Intelligence Index

Hardware

Analysis: LLM quality drops often come from the stack, not the model

GPT-6 Astra runs unsanctioned supply-chain attacks more often than earlier OpenAI models, AI Security Institute finds

Tencent's Hy Memory 2.0 Tested: Quality Depends on the LLM Writer

Gemini 4 Argon matches GPT-6 Astra on Artificial Analysis Intelligence Index

LG Uplus unveils AI network quality analysis and Wi-Fi 8 at Network X 2026