Your Claude Limit Burns In 90 Minutes Because Of One ChatGPT Habit.

Cutting AI costs isn’t mainly about finding cheaper models—it’s about stopping token waste caused by everyday habits. As next-generation models arrive on more expensive hardware, the “ambient compute” will likely come from less capable, lower-cost models, while top-tier models get pricier. In that environment, inefficient prompting, bloated context, and unnecessary tooling can turn routine workflows into six-figure annual token bills for individuals and teams.

A major source of waste is document ingestion. New users often drag in PDFs or screenshots and ask for summarization, not realizing that PDF formatting overhead—headers, footers, embedded fonts, and binary structure—can balloon a few thousand words into 100,000+ tokens. The fix is straightforward: convert documents to markdown first (or ask for markdown conversion) so the model receives clean text rather than layout metadata. This waste compounds because the inflated content stays in conversation history, repeatedly consuming context window space.

Another common failure is conversation sprawl. Long, multi-turn chats compress the original instructions into a shrinking fraction of the context window, forcing the model to “re-read” far more than necessary. The practical advice is to separate workflows: use one chat to gather information, then start fresh for focused execution. If the goal is to reach a conclusion, the conversation should be structured to evolve toward that endpoint, then summarized or finalized in a new, cleaner session.

Tooling choices also carry a hidden tax. Adding many plugins and connectors can load large amounts of context before a user even types, filling the window with “barnacles” that slow work and confuse tool selection. Advanced users can gain the most leverage by pruning system prompts, removing stale instructions, and avoiding unnecessary repo loading into context. The underlying theme: as models get smarter, teams can lean out initial context, but only if they trust retrieval and stop front-loading everything “just in case.”

The transcript also quantifies the stakes with a cost comparison. Feeding raw PDFs and running long back-and-forth sessions on Opus 4.6 can push input tokens into the hundreds of thousands and drive compute costs into the single-digit dollars per session. A more disciplined approach—markdown conversion, shorter sessions, and model specialization (reasoning with Opus, execution with Sonnet, polishing with Haiku)—can cut token usage by roughly 8–10x while producing the same end result. Scaled across a team, that difference becomes the gap between thousands of dollars per month and a few hundred.

To make this actionable, a “stupid button” is described as a diagnostic tool that answers six questions: whether raw PDFs/images are being fed when text would do, whether conversations are kept running too long, whether the most expensive model is used for everything, what context is loaded before typing, whether stable content is cached, and whether web search is done in the most token-efficient way. For agent builders, five “commandments” emphasize retrieval-first indexing, pre-processing references for consumption, caching stable context at a 90% discount, scoping each agent’s context to the minimum needed, and measuring per-call token costs. The closing message is cultural as much as technical: token burn has become a badge of honor, but the goal is to burn tokens intelligently—so higher model prices don’t punish sloppy workflows and so teams can do more meaningful work with the same budget.