Spend fewer tokens when you combine tools: cache, keep context short, and batch the work that can wait

The scene

You already wired mail, docs, an editor, or office apps to a model, and the bill is climbing faster than you expected. You want to cut waste without abandoning the tools.

Tools involved

What they did

1. Find the text you paste on every call: the role, the project rules, the style guide, or a fixed knowledge base. Most official APIs charge a lower cache-hit price when that same prefix comes back.
2. Compare official cache rates. Claude Sonnet 5.5: $2 / MTok base input, $0.10 / MTok cache hits, $2.50 for 5-minute cache writes, and $4 for 1-hour cache writes. OpenAI gpt-5.4-mini: $0.75 standard input and $0.075 cached input. Gemini 3.8 Flash lists a paid cache price of $0.075 per million tokens through 31 Dec 2026, plus an hourly storage price. DeepSeek flash off-peak cache-hit input is $0.003, versus $0.15 on a cache miss.
3. Attach only the files or messages the current step needs. Do not pour a whole inbox or repository into the context. Longer inputs cost more, and some models add a long-context surcharge — OpenAI’s official table lists separate long-context prices for some models.
4. If a job can wait a few hours — cleanup, translation, archiving — look at Batch. OpenAI lists Batch separately; gpt-5.4-mini Batch is $0.375 / $2.25, about half the standard rate. Anthropic’s official page says you can save 50% with batch processing.
5. When several apps are chained, pass a short structured result (a list, fields, a diff) to the next step instead of forwarding the previous long answer unchanged. A person should still review anything that will send mail, pay, or overwrite a file.

Copyable prompt

Here is the system prompt and the tools I call on every run. Point out which parts I should cache, which files I should stop attaching by default, and which jobs could wait for a batch window. Use only the official cache and batch prices I paste below. If I did not paste a price, say so and do not invent one.

Rewritten here for you to try; not a verbatim excerpt from elsewhere. What it asks：Give the model your standing prompt, the tools you always call, and the official cache or batch prices you pasted, and ask it what to cache, drop, or batch.

Things to watch

Cache has a write price and a lifetime (Anthropic lists 5-minute and 1-hour tiers); stored text does not stay at the hit rate forever. Google also lists an hourly cache storage price. Tool calls, web search, and code execution may be billed separately from these token rates. Figures were checked on 10 Nov 2026; if an official page changes later, use the new page.