Every prompt you send to Claude or ChatGPT costs tokens, and tokens cost money and latency. The good news is that most token waste isn’t in the model — it’s in the prompt. A few structural habits can cut usage by 30-60% without hurting output quality. Here’s what actually moves the needle.
Trim the system prompt
If you’re calling either model’s API, your system prompt runs on every single request. A bloated 500-word system prompt multiplied across thousands of calls adds up fast. Keep it under roughly 100 words and include only what the model actually needs to behave correctly — cut the hedging, the repeated caveats, and the examples that belong elsewhere.
Put reference material first, instructions last
When a prompt includes a long document, transcript, or dataset, place it above your actual question or instruction, not below it. Both models handle this ordering better, and it lets you reuse the same reference block across multiple questions without restating it. Wrapping sections in simple tags (like <document> or <instructions>) also helps the model separate context from task, which reduces the need for clarifying re-prompts.
Use fewer, better examples
Few-shot examples are one of the most common sources of silent bloat. Start with a single well-chosen example. Only add more if the output still misses the mark — and if you do add more, prune the earlier ones that no longer add information. Three redundant examples rarely teach the model more than one precise one.
Say the output format once, explicitly
Instead of letting the model guess and then re-prompting to fix formatting, state it upfront: “respond in three bullet points,” “return minified JSON, no explanation,” or “under 150 words.” Unformatted, pretty-printed JSON with indentation and whitespace can cost noticeably more tokens than a minified equivalent — and every round of “shorter, please” adds a full extra exchange to the conversation.
Keep conversations short and scoped
Every turn in a chat resends the full history, so long-running conversations get expensive fast — the model is re-reading the entire thread on every reply. Start a new conversation when the topic shifts, and periodically ask for a summary you can carry forward instead of scrolling back through dozens of turns.
Send only what’s relevant
Paste the specific section, function, or paragraph you’re working on rather than an entire file or webpage. If you’re working with scraped web content, strip navigation menus, ad markup, and boilerplate before it goes in the prompt — none of it helps the model, and all of it gets billed. For code, referencing a function by name or sending a diff is almost always cheaper than pasting the whole file.
Match the model to the task
Not every task needs the largest, most capable model. Simple summarization, formatting, classification, and rewriting tasks run well on smaller, faster models (like Claude Haiku or GPT’s mini variants) at a fraction of the cost. Reserve the flagship models for tasks that genuinely require deeper reasoning.
The takeaway
Token efficiency isn’t about writing less — it’s about writing precisely. A tight system prompt, well-ordered context, one good example instead of five, an explicit output format, and the right model for the job will typically cut token usage by a third or more, often