No one understands large models better from a consumer's perspective than I do.

1. For drawing, the top choice is the nano banana model (Free to use on Xianyu for 80 a year).
2. For videos, the top choice is seedance 2.0 (at least 2 dollars per minute, five videos are needed for effective results).
3. For engineering code, the top choices are Gpt 5.4 / Opus 4.6. Both are quite good, but the Opus ecosystem is better.
4. For understanding human nature, the top choice is Claude AI. I haven't seen any obvious hallucinations on Claude, and it understands human nature and politeness.
5. For AI rural outreach, the top choice is Doubao, exclusively for elderly family members, with basically no complaints.
6. For research papers, the top choice is Qwen. The models on crossreference with Qwen are truly effective.
7. For web front-end design, the top choice is Kimi2.5. The model has been trained with many good front-end UI designs, and can basically match Opus. If prompts are more precise, Kimi can slightly outperform Opus 4.6.
8. GLM / Minimax / Kiro—don't renew these anymore. If you want a brain, they are just slaves.