← sources

Does more thinking make LLMs write better? We tested every effort level across 7 model families. It helps Claude a lot and does nothing for Gemini

source: reddit (reddit:singularity)published 2026-09-29 18:01ingested 2026-09-30 02:41

retained text

We benchmark models on long-form YouTube scripts, and this month we ran every thinking/effort level each provider exposes. The chart plots writing score (Elo) against cost per script for a cross-section of models.

   
\- Thinking helps some families a lot: Claude Opus 5.5 gains +502 Elo from default to max, DeepSeek V4 Pro 0813 +463.  
   
\- GPT gains are real but smaller: GPT-6 Sol +150, GPT-5.6 Sol only +57 from default to ultra.  
   
\- Gemini, Grok, Qwen and Kimi don't benefit. Gemini 3.1 Pro on high thinking scores slightly below default, and Kimi K3's thinking mode drops 119 Elo.  
   
\- The top config, Opus 5.5 at max, is the only one that clearly beats our human-edited reference scripts (2846 vs 2663 Elo). It costs $3.43 and about 17 minutes per script. Opus 5.5 at xhigh matches the human scripts for $0.86.  
   
How to read the chart: writing score (Elo) against cost per script on a log scale. Each line is one model going from its lowest to its highest thinking setting, hollow markers are the default (no effort flag), and up-left is better.

The benchmark is 10 script tasks; each model drafts them 5 times, and three AI judges from three different labs score them blind using our rubric, which includes writing, tone, storytelling, and more.

[](https://www.reddit.com/submit/?source_id=t3_1wsw08f&composer_entry=crosspost_prompt)

document fetched 2026-09-30 00:33 · canonical https://reddit.com/r/singularity/comments/1wthafx/does_more_thinking_make_llms_write_better_we