8 ways to use fewer tokens (without hurting quality)

Token bills shrink from four levers: shorter inputs, shorter outputs, fewer calls, cheaper models. Everything below is one of the four.

1. Cut the prompt, not the instruction. Remove greetings, repeated context, and examples that don't change the answer. Trimming a 5,000-token document to 3,000 saves 2,000 input tokens per call — at 1,000 calls a month on GPT-6.1 Sol, that's $4.00 a month for free.

2. Cap the output. Output tokens cost 5× input tokens. Ask for the length you need (“in 5 bullets”, “under 150 words”) instead of accepting the model's default length.

3. Summarize before reasoning. For long documents, summarize to 20% first, then reason over the summary. A book-length input drops from 106,400 tokens to about 21,280 for every downstream call.

4. Cache what repeats. System prompts and reference documents are the same every call. Prompt caching (where the provider offers it) bills repeated prefixes at a steep discount.

5. Use structured output. JSON or a fixed schema stops the model from padding answers with prose you then have to parse — and pay for.

6. Batch small jobs. Ten separate calls each repeat the system prompt; one batched call pays it once.

7. Route by difficulty. Send classification, extraction and first drafts to a mini/flash model; escalate only the calls where the flagship's answer is visibly better. Most workloads are 80% routine.

8. Measure before optimizing. Run a week of real prompts through the token counter and sort by tokens × calls. The top 3 prompts are usually half the bill — fix those first.