90% of People Are Wasting 'Tokens'!
TL;DR · AI Summary
Karpathy points out that 90% of the AI coding bill is spent on unnecessary context, and optimizing context usage and routing strategies can significantly reduce costs.
Key Takeaways
- 90% of the AI coding bill is spent on unnecessary context.
- Use multi-model routing strategies, such as using Kimi 2.6 as the main model and
- Optimizing context and routing strategies can significantly reduce AI coding cos
Outline
Jump quickly between sections.
Karpathy points out that 90% of the AI coding bill is spent on unnecessary context.
Lists some common waste behaviors, such as automatically loading large files and using expensive models for simple tasks.
Proposes some optimization strategies, such as using multi-model routing and creating SKILL.md files.
Optimizing context and routing strategies can significantly reduce AI coding costs.
Mindmap
See how the topics connect at a glance.
查看大纲文本（无障碍 / 无 JS 友好）
- AI 编码优化
- 常见浪费行为
- 自动加载文件
- 使用昂贵模型
- 优化策略
- 多模型路由
- 创建 SKILL.md
Highlights
Key sentences worth saving and sharing.
Karpathy points out that 90% of the AI coding bill is spent on unnecessary context.
Use multi-model routing strategies, such as using Kimi 2.6 as the main model and Opus only for critical tasks.
Optimizing context and routing strategies can significantly reduce AI coding costs.
Berryxia.AI on X: "90% of People Are Wasting Tokens for Nothing!"
Seeing a post from @DeRonin_ that was shared by Andrej Karpathy, I found it incredibly insightful for those who code daily!
Karpathy bluntly states: "90% of your AI coding bill is actually spent on sending unnecessary context."
He then lists 10" / X
Don’t miss what’s happening
Show translation
90% of People Are Wasting Tokens for Nothing! Seeing
share a view from Andrej Karpathy, I found it incredibly insightful for those who code daily! Karpathy bluntly states: "90% of your AI coding bill is actually spent on sending unnecessary context." He then lists 10 common wasteful practices that senior engineers have stopped doing, and I’ll highlight a few of the most common ones:
For example, automatically loading 50 files to modify just 30 lines of code, resulting in a $1.20 token cost that you won’t even read.
Or using Opus for linting, formatting, and renaming tasks, which can be done with Haiku for just two cents, making it 30 times cheaper.
Another example is agents retrying by resending the entire repository each time, increasing costs fivefold.
Now everyone defaults to Sonnet, but Kimi 2.6 performs similarly in most coding tasks at only 1/6th the price.
Also, throwing all files into the prompt “just in case,” when 80,000 tokens could be reduced to just 3,000.
Rebuilding knowledge from scratch for each session instead of writing a SKILL.md file to save money.
Finally, he says that truly saving money while getting things done involves tightly managing context, enabling stable prefixes and prompt caching, using multi-model routing (with Kimi 2.6 as the main model and Opus reserved for 10% critical tasks), turning repetitive work into SKILL.md files, and profiling tool calls before optimizing prompts.
In short, in 12 months, the difference between a developer spending $200 and $4,000 per month isn’t about who has better skills, but who understands context and routing better. It’s worth every heavy AI coder taking a good look.
Quote
Ronin
@DeRonin_
·
7h
Andrej Karpathy: "90% of your AI coding bill is paying for context you didn't need to send" Here are 10 things senior AI engineers stopped wasting tokens on: 1. Auto-context loading 50 files for a 30-line fix: $1.20/turn for tokens you'll never read. 80% input waste, every x.com/DeRonin_/statu…
·
2
1
5
8