Claims of "90% token savings" from DIY tricks are probably overstated

A Reddit user pushed back on claims that individuals can build a prompt or tool that cuts token (context) usage by 90%. The argument: companies like OpenAI, Anthropic, and Google are already spending millions or billions to make token management as efficient as possible, so it's unlikely someone could easily match or beat that from home with a few clever prompts or tactics. It's compared to claiming you cut a car's gas consumption by 90% with a single home idea.

The core issue is that managing context is a limitation built into how large language models (LLMs) work, and reducing token use while keeping speed requires serious engineering. Every major AI provider is currently racing to solve this problem.