The video addresses the common frustration of hitting usage limits in AI tools like OpenAI’s Codex and Anthropic’s Claude, especially for users relying on them for serious work. Recently, both companies have reduced their usage limits, making it even easier to run out of tokens. The speaker analyzed their own usage logs and discovered that the majority of token consumption—around 95%—comes not from the AI generating responses or user input, but from the AI rereading the entire chat history repeatedly. This inefficiency is due to how these models process conversations, requiring them to reread all previous messages each time a new input is sent, which quickly adds up in token usage.
To combat this, the speaker recommends several habits to extend usage limits without paying more. First, avoid returning to old chats after an hour; instead, start a new chat with a summary or “handover note” to reduce rereading. Second, keep one job per chat and provide the full instructions upfront to minimize back-and-forth rereading. Third, choose moderate model effort levels and avoid using the fastest or most resource-intensive modes unless absolutely necessary. Fourth, use expensive models like Claude’s Fable or Codex’s Astra as orchestrators that delegate tasks to cheaper subagents, which can execute work more efficiently and cost-effectively. Lastly, regularly clean up your chat context by removing unused plugins, instructions, or tools that unnecessarily bloat the token count.
The video also explains how the AI’s context window grows over time, with initial instructions and tools often taking tens of thousands of tokens before any user input. As chats get longer, every new message requires rereading an increasingly large context, which drastically increases token consumption. The speaker shares personal data showing how a single long-running chat can consume a significant portion of monthly usage. They emphasize the importance of starting fresh chats after completing tasks and consolidating workflows into fewer, more comprehensive messages to reduce overhead.
Additionally, the speaker warns against common misconceptions that don’t save tokens, such as typing shorter prompts or waiting for the chat to compact automatically. They highlight that certain file types, like PDFs or images, are more expensive for the AI to process compared to text-based formats like markdown. They also caution that upgrading to more expensive subscription plans with faster modes can actually accelerate token consumption, leading to quicker exhaustion of limits. Switching models mid-chat also triggers rereading of the entire context, so consistency is advised.
In conclusion, the video stresses that AI usage limits are likely to continue shrinking as companies prepare for IPOs and investors push for profitability. Users must adopt smarter habits to maximize their token usage efficiently. The speaker provides a prompt tool to analyze personal usage logs and identify which habits will yield the greatest savings. By following the five key habits—avoiding returning to old chats, limiting one job per chat, choosing moderate effort levels, delegating tasks to cheaper subagents, and cleaning up context—users can significantly extend their AI usage without additional cost.
Useful Links
