Spend less. Get the same answers.

To reduce LLM costs, send fewer tokens to cheaper models, and stop paying full price for work you repeat. In practice that means routing easy tasks to small models, caching the parts of your prompt that never change, batching anything that can wait and capping what the model writes.

None of it requires a weaker model for the hard parts, and every discount below is documented on the vendors’ own pages. Prices were checked on September 22, 2026. They change often, so confirm them before you plan a budget.

01How to reduce LLM costs: the eight levers

Every lever below either lowers the price per token or cuts the number of tokens.

Watch the input you never typed. In a chat or agent loop, the whole conversation is resent on every turn. Tool definitions ride along too: Anthropic’s computer use toolset alone adds about 4,500 input tokens per request. Anthropic also notes that its newer tokenizer produces about 30% more tokens for the same text than older Claude models. Compare cost per task, not price per token.

For scale, Anthropic’s own worked example puts 10,000 support conversations of about 3,700 tokens each at roughly $37 on Claude Haiku 4.5. New to tokens? Start with tokens and context windows.

02Use the smallest model that does the job

This is the biggest lever. Anthropic’s docs make the point directly: “you can sometimes improve latency and cost more easily by selecting a different model.”

Standard API prices in US dollars per million tokens, as of September 22, 2026. Gemini 3.8 Flash is scheduled to rise to $1.50 and $7.50 on January 1, 2027. Gemini 3.1 Pro charges more for prompts over 200,000 tokens. For the fuller picture, including cached and batch prices, see our side-by-side API price table.

To route well, collect 20 to 50 real requests for one task, with answers you consider good. Run them through the cheapest plausible model first and grade the results. Move up a tier only for the cases it fails. A cheap model can also sort incoming requests, sending only the hard ones to a larger model.

Rerun the set when new models ship. Our guide to choosing an AI model covers the trade-offs.

03Cache the part of the prompt that repeats

Most prompts start the same way every time: instructions, tool definitions, reference documents. Prompt caching stores the processed version of that shared start. Later requests that begin with identical text read it at a steep discount.

Anthropic charges extra to write the cache: 1.25 times the input price for five minutes, twice for an hour. By its own math, the five-minute cache pays off after one reuse and the hour-long one after two.

Two rules make caching work. First, put stable content first and anything that changes last, as both OpenAI and Google advise. A timestamp at the top of your system prompt breaks the cache on every call. Second, watch the cached token count that each response reports. If it stays at zero, something near the top is changing.

1. System instructions     same on every call
2. Tool definitions        same on every call
3. Reference documents     change rarely
4. Conversation so far     grows each turn
5. The new request         different every time

04Batch what can wait, and skip paid speed

If a job does not need an answer in seconds, do not pay for seconds. Anthropic, OpenAI and Google all run batch APIs: you submit many requests at once and collect the results later, at half price.

Good batch jobs include evaluations, classifying a backlog, embedding a document collection and nightly reports. Discounts stack: Anthropic says its caching discounts apply on top of the batch discount. OpenAI’s Flex processing, in beta on selected models, charges batch prices for ordinary requests in exchange for slower responses and “occasional resource unavailability.”

Premium speed works the other way. Fast modes at Anthropic and OpenAI cost twice the standard rate, and OpenAI renamed its Priority processing to Fast mode on July 30, 2026. US-only processing on Anthropic’s API adds 10%. Pay for these only when a person is waiting or a contract requires it.

05Send fewer tokens in and out

Trim the context

Every token in the prompt is billed on every call. Delete instructions that no longer change the output. Turn on only the tools a task needs. Retrieve the relevant passages instead of pasting whole documents: Anthropic estimates a research paper PDF at about 125,000 tokens. Summarize or restart long chats instead of carrying the full history.

Watch for price thresholds too. Gemini 3.1 Pro doubles its input price above 200,000 tokens, while Anthropic’s Claude 4.6 and later models charge the same rate across their full million-token window.

Cap the output

Output is the expensive side. Set a maximum length on every call, and ask for exactly the shape you need: a JSON object, a table, three bullets. For code, ask for a diff or the changed function rather than the whole file.

Answer in at most 5 bullet points, each under 20 words.
No introduction, no summary, no restating the question.
If you need more room to be accurate, reply "Needs more detail" instead of writing more.

Set reasoning effort per task

Reasoning models think before they answer, and OpenAI’s docs say that hidden reasoning is “billed as output tokens.” OpenAI and Anthropic both let you turn it down. OpenAI’s reasoning effort setting has levels from none to max, depending on the model, and “lower effort favors speed and lower token usage.” On Anthropic’s models, “higher effort elicits more thinking.”

Use low effort for extraction, formatting and simple questions, and save high effort for planning, debugging and hard analysis. Do not add “think step by step” to a reasoning model: a 2025 Wharton study found it gave marginal gains at best while adding time and tokens. See when reasoning models are worth it.

06Stop paying twice, and pick the right meter

Reuse results

Provider caching discounts repeated input. Your own cache can skip the call entirely. Store answers to repeated questions, classifications and lookups. Re-embed only the documents that changed, and remove duplicates before a batch job.

Be careful with caches that match by meaning: “cancel my order” and “cancel my subscription” sound alike but need different answers.

Subscription or API

A plan gives one person a usage allowance for a flat fee. The API bills every token, with no allowance and no ceiling.

If one person uses AI heavily every day, estimate what a typical week would cost at API prices and compare it with the plan. Some plans also bend at the limit: Claude’s paid plans let you keep working with usage credits billed at standard API rates.

07A monthly cost review

08FAQ

Next, see whether running models on your own machine makes sense for you, or compare open-weight and closed models.