UGI leaderboard: Gemini 4 Argon writes best at low thinking, least censored at high
Eissa_Cozorav · reddit · 2026-10-18
Per the UGI Leaderboard on Hugging Face, google/gemini-4-argon's behavior shifts sharply with its thinking level: thinkinglevel=low scores highest on Writing (79.71) and is best as a story writer, while thinkinglevel=high jumps to UGI 49.92 (W/10 1.8), making it the least censored setting. Medium splits the difference (UGI 33.33, Writing 79.04). Compared to predecessors: gemini-3.8-flash hits UGI 44.9 and anthropic/claude-fable-5 leads at 55.92. Practical takeaway: low thinking for prose quality, high thinking for fewer refusals.
More from Models
- Model repeatedly claims to be ChatGPT once OpenRouter default system prompt removed — scaling01 · 2026-10-18
- New model repeatedly identifies itself as ChatGPT in hands-on test — scaling01 · 2026-10-18
- GLM 5.3 finds arbitrary-write vulnerability in Apollo 11 guidance code — moyix · 2026-10-18
- Google AI flags its own generated character as a celebrity mid-project — noisywan · 2026-10-18
- Grok's "complex age dynamics" phrasing for historical abuse draws flak — eigenrobot · 2026-10-18
- Dev revisits Opus 4.6 project with Opus 5.5, finds it spotting old bugs unprompted — rms80 · 2026-10-18