How to reduce your LLM API costs

Before you change anything, split last month's bill by model, by key, and by tool. Most bills have one or two lines that matter: an agent that loops, a script that sends a whole file every time, or a top model answering questions a small one would get right.

Two things surprise people. Output tokens cost far more than input tokens: GPT-5.5 lists $5 per million in and $30 per million out. And long chats send the whole conversation again with every message, so the tenth message in a chat costs much more than the first.

Not every request needs the top model. A quick question, a rename, or a short summary is often answered just as well by a cheaper one. List prices per million tokens, input then output, show how wide the gap is:

You can do this by hand: pick a cheap default and switch up when a task is hard. A router does it per request. FUSR.ai's fusr/auto sorts each request by task and sends quick questions to the cheapest connected model first. Turn on cheapest-good-enough and every task goes to the cheapest model rated good enough for it. How the AI router works.

Routing is not magic, and sometimes it saves nothing:

• Your work is already on a cheap model.
• Almost every request is hard, so it all belongs on the top model anyway.
• Switching models in the middle of a chat throws away the prompt cache, which can cost more than the cheaper model saves.

A well chosen pair, one strong model and one cheap one, gets most people a long way. Caching, caps, and trimming what you send often matter as much as clever routing.

When the start of a request is the same as a recent one (a system prompt, a file, the earlier part of a chat), providers can read it from a cache for much less than full price. Anthropic charges a tenth of the input price for cache reads, and a twentieth on Sonnet 5.5 and Opus 5.5, but a quarter more to write the cache. OpenAI and xAI cache on their own: GPT-5.5 lists cached input at $0.50 against $5, and Grok 4.7 at $0.50 against $2.

Habits that break caching:

• Switching models mid chat. Each model has its own cache.
• Putting something that changes, like the time, at the top of a system prompt.
• Long pauses. Cache entries expire after a few minutes without use.

FUSR.ai marks Anthropic requests for caching, keeps a chat on the model whose cache is warm, and shows what caching saved in your report, from your real token counts.

One runaway script or agent loop can spend a month's budget overnight. Provider consoles offer budgets and alerts for an organization or project. A cap per key is tighter: each tool or teammate gets its own limit, and hitting it stops only that key.

FUSR.ai Free has one monthly cap on your FUSR.ai key. Pro and Team set daily and monthly caps on every key, with an alert at the dollar amount you choose.

OpenAI and Anthropic both offer a batch API for jobs that can wait up to a day, such as tagging a backlog or running evals, at half the normal price. FUSR.ai does not route batch jobs, so send those to the provider directly.

• Start a new chat when the topic changes, instead of dragging the old context along.
• Send the part of a file that matters, not the whole repository.
• Set a max output length for jobs that only need a short answer.
• Keep system prompts short and stable, so they cache well.

GPT-5.5 is the priciest of the default models, at $30 per million output tokens. Keep it for the work it is best at, let quick questions go to a cheaper model, keep the start of your prompts stable so OpenAI's automatic caching kicks in, and set a max output length. The same key can reach Claude and xAI models through FUSR.ai if you connect them.