65-Day Analysis of 43K Claude Code Calls Finds Thinking Budgets Silently Slashed
IcyEase · reddit · 2026-09-20
- 一份覆盖 65 天、43000+ 次 Claude Code 调用的分析发现:39% 的 Fable 5 调用拿到零思考 token,中位数仅 123 tokens,而官方基准测试使用 16K-128K。
- 模型每思考 token 的得分在 128K 处仍在上升,说明能力存在但未交付;8 月思考预算较 7 月下降 18-50%,8 月 22 日前后约一周中位数思考量直接归零。
- 作者指控 Anthropic 一边售卖「完整模型访问」,一边悄悄下调推理配置;由于模型输出非确定性,用户往往归咎于自己的提示词而非静默降配。发帖人认为这才是付费用户该关注的实质——为所付价格对应的真实能力,而非每周用量限额。
Related event: Study Claims Claude Users' Reasoning Compute Is Secretly Throttled(2 posts)→
More from Companies & People
- Bay Area AI lab engineer: curiosity compounds better than a career plan — itsOmSarraf_ · 2026-11-12
- Microsoft and Ohio State announce five-year AI healthcare partnership — luisdans · 2026-11-12
- OpenAI, Anthropic and DeepMind job posts signal PMs must write evals and code — aakashgupta · 2026-11-12
- PowerShell creator Jeffrey Snover at Ignite: AI revolution is an extinction event for those who don't adapt — dfinke · 2026-11-12
- OpenAI CFO Sarah Friar says OpenAI has 100 million subscribers — morqon · 2026-11-12
- Rox raises $270M Series B at $1.3B valuation, 12 months after launch — Konstantine · 2026-11-12