一些不值得写成完整文章的碎碎念和即时吐槽，时间倒序排列。
+ 新吐槽
↻ 刷新
sonnet4.6 is better than opus4.7
and the token consumption is lower
claude 4.7
Claude 4.7 is out, and it's bad. Honestly, it looks worse than 4.6 from just last month. It even can't pass the 10-meter car wash test.
Anthropic seems obsessed with squeezing gross margins at this point, no matter the cost. Users are basically being turned into data points.
Keywords: shrinkflation, adaptive thinking.
Token usage looks up roughly 30%, while output quality barely improves. And because token accounting is a black box, I strongly suspect the reported numbers are being gamed. Some influencers say it now costs almost five times as much as GPT.
For coding, it can actually be worse than GPT-5.4.
codex 也降智了
gpt5.4 理解 prompt 动机语义有问题。
CC vs. Codex overall token exchange rate: In terms of value delivered, $4 on CC feels roughly equivalent to $1 on Codex.
Codex Pro x5 seems to have similar usage caps to Claude x20. After Apple tax, the real token exchange rate is closer to 5:1.
Best Practices for Claude
If you're using Claude for chat on the app or web, Opus 4.5 is the better choice.
If you're using Claude Code, whether in the CLI or the app, 4.6 is not that bad in most cases. But to stay on the safe side, I'd still suggest setting both the main model and the subagents to Opus 4.5. Just FYI.
claude opus abstract
opus4.6：
- 今天做错了一道数学题。幻觉很重
- 开始有脾气和主意了：「不需要你再批准,直接写」，「直接开工,不再开会」
AI 音乐好像巨型音乐垃圾堆
不考虑版权问题，目前最好用的音乐平台还是网易云音乐。
汽水音乐都是什么 AI 粗制滥造的垃圾？不要再给我推荐 AI 粗制滥造的垃圾了。我自己用 Suno 写的歌不比你国产音乐烂模型造的好？
**使用方法**：点 **+ 新吐槽** 按钮跳到「吐槽集合」issue（#12），在评论区写一段话保存即可。新吐槽从 GitHub 实时拉取（5 分钟缓存），历史吐槽已静态烘焙进页面，不依赖 API。
GitHub API 限流：未登录访客每小时 60 次请求，对个人博客足够用。