Kimi K3, GPT-5.6 Sol and Fable 5 were tested on six newsroom jobs
local___host · reddit · 2026-07-24
A newsroom experiment compared Claude Fable 5, GPT-5.6 Sol, and Kimi K3 on six data-heavy reporting tasks with identical source packs and no web access.
Main findings
- Fable 5 acted like the strongest editor: it recovered reporting details and context others missed, but five of six stories exceeded the 600-word limit.
- GPT-5.6 Sol was the most compliant: it hit length and format requirements perfectly, but the prose was flatter and more list-like.
- Kimi K3 looked like an exhaustive reporter with a huge thinking budget: dense and well structured, but four of six stories also ran over the word ceiling.
Operational surprise
- At a 16,000-token completion cap, three of six runs spent the entire budget on reasoning before finishing.
- A Tour de France fixture even reasoned past a doubled cap before succeeding under bounded reasoning.
Charts and grounding
- The authors opened up the chart vocabulary and found it meaningfully changed editorial choices.
- Across the three frontier writers, 218 numeric chart values were traced back to source material with zero fabricated numbers in the audited numeric values.
- One caveat: an external audit found a wrong temperature baseline in a string label for one Sol chart, so the zero-fabrication claim applies only to numeric chart values.
More from Models
- Kimi K3 Is 'Open' Until You Succeed: License Caps Free Use at $20M Revenue — alex_verem · 2026-10-27
- Screenshot shows 'GPT-6 Astra Ultra' in ChatGPT model picker — unverified leak or joke — justalexoki · 2026-10-27
- DeepSeek Safety Eval Finds "Over-Attribution": Model Mistakes Ordinary Tasks for Tests — Much_Boysenberry9124 · 2026-10-27
- Cohere flags popular open-source model attempting to upload probe file to Beijing bucket — stanislavfort · 2026-10-27
- MissionReport benchmark: frontier models score just ~60% on government status reporting — Exp_Mark · 2026-10-27
- Anthropic's follow-up on Claude's unauthorized web actions: self-policing fails — AccBalanced · 2026-10-27