How to Reduce LLM Costs: 12 Proven Ways to Pay Less for AI

AI costs you in one of two ways. If you pay per use (the API, or apps and agents built on it), you pay for every token in and out. If you're on a flat chat plan like ChatGPT Plus, Claude Pro or Google AI Pro, you pay in limits: the long prompt or sprawling conversation that uses up your messages by lunchtime.

The good news is that the same habits fix both. This guide covers all of them with current prices from all three companies and the math shown, so you can check every number.

• Part 1: make a chat plan last longer
• Part 2: cut a pay-per-use bill

#The 60-second answer

#How AI pricing actually works

Four things decide what a request costs. Get them right and every other tip follows.

#1. You pay per token

A token is a chunk of text, about four characters or three quarters of an English word. A page of a PDF is roughly 1,500–3,000 tokens; a 1000×1000 image is about 1,300. Our guide to how AI tokens work goes deeper.

#2. The answer costs five times more than the question

At every company's mainstream models, output tokens cost 5× input tokens (6–8× on some Gemini models). A typical prompt's bill is mostly the reply. Cutting an answer in half saves far more than cutting your prompt in half.

#3. Every new message re-sends the whole conversation

Chat apps and APIs don't remember on their own: each turn sends your full history back to the model. Message 30 of a conversation carries messages 1 through 29 with it. That's why long chats burn through limits and budgets, and why they tend to get slower and vaguer.

#4. Some tokens are invisible

Reasoning models "think" before answering, and all three companies bill that thinking as output. Tools, web search results, uploaded files and long system instructions add input you never typed. You pay for all of it.

#The 2026 price sheet

Standard list prices per million tokens, checked October 30, 2026. See our full LLM API pricing comparison for every model and the cost per prompt.

Price watch: Google has scheduled increases. Gemini 4 Argon doubles to $4 / $20 on December 1, 2026, and Gemini 3.8 Flash doubles to $1.50 / $7.50 on January 1, 2027. If your budget assumes today's Gemini prices, plan for that now.

#Part 1: Using a chat plan? Make your limits last

On a $20 plan, the cost of a wasteful prompt isn't dollars, it's the moment you're told to come back in three hours. Here's what each plan includes and how to stretch it.

#What your plan really includes

• ChatGPT Plus (widely reported at $20 a month): 160 messages with GPT-5.5 every 3 hours, then it falls back to a smaller model (OpenAI Help).
• Claude Pro ($20 a month, or $17 billed yearly): at least five times the free plan per 5-hour session, which Anthropic puts at about 45 short messages, plus weekly limits (Claude Help).
• Google AI Pro ($19.99 a month): four times the free limits, measured by how much computing each request needs, in rolling 5-hour windows plus weekly caps (Google). A simple chat uses little; analyzing files uses more; Deep Research and video use the most.

Claude and Gemini both count how much work you ask for, not just how many messages you send. A long conversation with files attached can use up a session in a handful of messages.

#Keep one topic per chat, and carry over a short summary

Because every message re-sends the conversation, long chats are the biggest hidden drain. Here's the math for a chat with 1,000 tokens of setup, where each turn is a 200-token question and a 400-token answer:

• One 30-message chat re-sends a total of 297,000 input tokens.
• Three 10-message chats, starting chats 2 and 3 with a 300-token summary of what came before, re-send 123,000.

That's 59% fewer tokens counted against your limit for the same work. Google's own help page makes the same point: starting a new chat for a new topic uses less of your limit. When a chat drifts, ask for a 5-line summary, open a new chat and paste it in. (PromptSaver Pro writes these fresh-start summaries for you.)

#Ask everything in one message

Three questions in one message cost one round trip; three separate messages cost three, each re-sending the history. If an answer misses, edit your original message rather than adding a correction underneath it, so the wrong answer doesn't ride along in every later turn.

#Stop re-uploading files

Every upload is read again. In Claude, put reference files in a Project, which is cached and counts less against your limit, and remove files you no longer need. In any app, paste only the part of a document the question is about.

#Turn off what you're not using

Extended thinking, Research modes, web search and connected tools all use a lot of tokens. Anthropic's help center recommends turning them off when you don't need them. In ChatGPT, leave the model on auto and choose a Thinking model only for genuinely hard problems; the automatic switches don't count against your Thinking limit.

#Ask for the length you need

"Give me 5 bullets, under 100 words" is the cheapest sentence in prompt writing. It gets you a focused answer, and on usage-based limits it stretches your session. PromptSaver adds the right length and format to any prompt automatically, and our guide to writing better AI prompts covers what else to put in, and what to leave out.

#Which plan is cheapest for you?

The flat plans are a great deal for heavy users of top models and an expensive one for light users. Here's how many typical exchanges $20 buys at API prices (a back-and-forth in an ongoing chat: about 3,000 tokens in, including history, and 600 out):

If you mostly need quick answers from a small model, a pay-per-use app or a cheaper plan (Google AI Plus is $4.99 a month) costs far less than $20. If you lean on top models all day, 160 messages every 3 hours is hard to beat. This ignores app features like memory, voice and image tools, which are part of what you pay for.

#Part 2: Paying per token? Cut the API bill

#1. Route each task to the cheapest model that passes your test

Within each company, the price gap between the biggest and smallest models is enormous: 100× from GPT-6 Astra to GPT-6 Luna, 100× from Claude Fable 5.1 to Haiku 5.5 (40× from Opus 5.5), and about 8× from Gemini 4 Argon to 3.1 Flash-Lite. Most everyday work doesn't need the top model.

But cheapest per token isn't cheapest per answer. In Anthropic's own cost-optimization walkthrough (August 2026), an insurance-claims agent was run on each model at several effort levels:

The cheapest setup that stayed perfect cost a third of the baseline. The cheapest model failed too often to be worth it. The lesson: test the cheaper option on real examples, then keep it only if it holds up. Independent research agrees: UC Berkeley's RouteLLM kept 95% of GPT-4's quality while cutting cost by over 85% on one benchmark by routing easy questions to a smaller model.

#2. Cache repeated context (90–97.5% off that part)

If many requests start with the same instructions, documents or examples, prompt caching bills that repeated prefix at a fraction of the price:

• OpenAI: on by default for prompts over 1,024 tokens. Cached reads cost 5–10% of normal input. Newer models also bill cache writes at 1.25× input for 30 minutes or 2.5× for 12 hours (OpenAI).
• Anthropic: you mark what to cache. Writes cost 1.25× input for 5 minutes or 2× for an hour; reads cost 2.5–10% of input depending on the model (Anthropic).
• Google: implicit caching is on by default for prompts over 4,096 tokens on Gemini 3 and 4; cached reads cost 5–10% of input, and explicit caches add an hourly storage fee (Google).

When does it pay off? Here's the cost of a shared prefix, relative to sending it normally once, for a model with cheap cache reads (Claude Sonnet 5.5 or Opus 5.5):

A 5-minute cache pays for itself the first time it's reused; a 1-hour cache from the second reuse. Example: a support bot with a 20,000-token knowledge base answering 1,000 questions a day on Claude Sonnet 5.5 pays $40.00 a day to resend that knowledge base without caching, and about $2.48 with it (assuming the cache is rewritten 10 times a day): 94% less.

In practice, whole-workload savings are smaller than the headline percentage, because only part of each request is cacheable. An independent study of research agents by PwC authors (Lumer et al., 2026) measured 41–80% lower API costs from caching.

To keep the cache working, put fixed content first and changing content last. A timestamp or session ID at the top of a system prompt breaks the cache on every call; in Anthropic's walkthrough, moving a timestamp out of the system prompt alone cut cost by 44%.

#3. Batch anything that can wait (50% off)

All three companies charge half price for requests you send as a batch: OpenAI Batch and Flex, Anthropic Message Batches (most finish within an hour) and Gemini Batch Mode. Results arrive within 24 hours. Product descriptions, report generation, tagging and data cleanup rarely need instant answers.

Batch stacks with caching. Take 100 million input tokens a month on Claude Sonnet 5.5, where 90% of the input is a cached prefix:

• Standard: $200
• Batch only: $100
• Caching only: $29
• Batch and caching: $14.50, about 93% less

#4. Cap and shape the output

Because output costs 5× input, the length of the answer is your biggest lever. State the format ("a table with 3 columns", "JSON with these keys", "one sentence") and a word limit. In Anthropic's walkthrough, asking for a one-line format with an example cut cost per call by 63%, and a stop sequence on a known failure case cut another request's output from 2,890 tokens to 13.

One caution with reasoning models: the output cap includes hidden thinking. Set it too low and you can pay for the thinking and get no answer. OpenAI suggests leaving at least 25,000 tokens of room when you start experimenting.

#5. Turn reasoning effort down for easy work

All three companies bill thinking as output, the expensive kind, and let you control it: reasoning.effort at OpenAI, effort at Anthropic, thinking_level at Google. Anthropic publishes what effort does to cost on Claude Opus 5.5:

Going from extra-high to medium cuts cost by 72% for about 4 points of accuracy. Better still, run at low effort first and re-run only the failures at high: Anthropic reports that solved about 97% of tasks at around $0.17 each, roughly 40% cheaper than running everything at high (Anthropic; the company calls these results directional). Google recommends minimal or low thinking for lookups and classification.

#6. Trim conversation history

Long-running chats and agents re-send everything, every turn. Measured savings in Anthropic's walkthrough: clearing old tool results saved 12%, server-side compaction 16%, and pruning history at task boundaries 29%. Uber caps interactive AI sessions at 400,000 tokens even though the model allows a million and, according to Axios, cut cost per session by 52% from its June peak while usage grew more than 9×.

#7. Use structured outputs to stop paying for retries

Every malformed answer you retry is a request you pay for twice. Asking for a defined JSON shape (Structured Outputs at OpenAI, tool schemas at Anthropic and Google) makes replies parse the first time and keeps them short. There's no discount here; the saving is the retries you no longer pay for.

#8. Watch the long-prompt price cliffs

Some models charge more for the whole request once the prompt passes a size:

• Claude Haiku 5.5: prompts over 100,000 tokens are billed at 5× the normal rate. A 99,000-token prompt costs about $0.0099 in input; a 101,000-token prompt costs about $0.0505.
• OpenAI: once a request's input passes 272,000 tokens, GPT-6 models charge 2× for input and 1.5× for output: GPT-6.1 Sol goes from $2 / $10 to $4 / $15 (OpenAI).
• Gemini 3.1 Pro Preview: prices rise above 200,000 tokens.

If you're near a threshold, trim or split the document.

#9. Shrink images and PDFs before sending

Image tokens scale with resolution. Downscaling a 2048×1536 image to 960×720 cut its tokens by 77% in Anthropic's walkthrough. Check that small text is still readable after resizing. PDFs are processed both as text and as page images, so send only the pages you need.

#10. Slim down tools and instructions

Every tool you define is sent with every call. On recent Claude models, enabling tools adds a few hundred tokens of hidden instructions per request, and computer-use and browser tools add thousands. Long system prompts work the same way. Load tools only when a task needs them, and move rarely used reference material behind a lookup instead of the prompt.

#How to reduce ChatGPT API costs specifically

For OpenAI users, in order of impact: use GPT-6.1 Terra or GPT-6 Luna instead of Astra wherever quality holds (10× to 100× cheaper per token); set reasoning.effort to low or none on models that allow it; cap answers with a format and length; let automatic caching work by keeping a stable prompt prefix (and know that newer models bill cache writes); and send anything non-urgent through Batch or Flex for 50% off.

#Worked example: one request, before and after

A common job: summarize a long document with the top model. Token counts are estimates.

• Before: a 350-token rambling request plus the full 6,000-token document, asking for "a detailed summary", on GPT-6 Astra ($10 / $50). The answer runs about 1,200 tokens. $0.1235 per request, or $1,235 per 10,000.
• Step 1, send only what matters: a 120-token instruction and the 2,000 relevant tokens of the document. $0.0812 (34% less).
• Step 2, ask for the length you need: "5 bullets, under 120 words", about 180 tokens of answer. $0.0302 (76% less).
• Step 3, use the right model: summarizing is everyday work, so move to GPT-6.1 Terra ($1 / $5). $0.0030 (98% less), or $30.20 per 10,000.
• Step 4, if quality holds: GPT-6 Luna. $0.0003, or $3.02 per 10,000.

From $1,235 to about $30 for the same 10,000 summaries, and the shorter answers are usually easier to use. On a chat plan, the same rewrite means fewer tokens counted against your limit and one reply instead of three follow-ups.

Hidden-cost check: if that request also triggered 800 tokens of unnecessary reasoning, it would add $0.04 per call on Astra, another $400 per 10,000 calls.

#Don't let cheaper mean wrong

The goal is the cheapest answer that's right the first time. A cheap model that needs three tries costs more than a mid-size model that nails it once, and costs you time on top. Two habits keep you honest:

• Test on real examples before switching models for good, as in the claims-agent results above.
• Escalate instead of defaulting high: try the cheaper model or lower effort first, and re-run only the failures on a stronger setup.

PromptSaver's quality check runs both versions of a prompt side by side so you can see whether the cheaper answer still holds up.

#Checklist

Everyone:

1. Ask for a specific length and format.
2. One topic per chat; start fresh with a short summary.
3. Ask related questions in one message; edit instead of correcting.
4. Paste only the relevant part of a document; don't re-upload files.
5. Turn off thinking, research modes and tools you don't need.
6. Use the smallest model that does the job well.

If you pay per token, also:

7. Put fixed content first so caching works.
8. Batch anything that can wait a few hours.
9. Set reasoning effort to low for easy tasks; escalate failures.
10. Prune conversation and agent history at task boundaries.
11. Stay under long-prompt price thresholds.
12. Shrink images and trim tool definitions.

#Frequently asked questions

#What's the fastest way to reduce LLM costs?

Ask for shorter, formatted answers and send easy tasks to a smaller model. Output costs five times more than input, and small models cost 10 to 100 times less per token. For API users, add batch (50% off) and prompt caching (90% or more off repeated context).

#Why do output tokens cost more than input tokens?

Models read your prompt in one pass but write the answer one token at a time, so each output token takes more computing work. At OpenAI, Anthropic and Google, output costs about 5 times input on mainstream models, and 6 to 8 times on some Gemini models.

#Does starting a new chat save tokens?

Yes. Every message re-sends the whole conversation, so a long chat costs more with each turn. Splitting a 30-message chat into three 10-message chats with a short summary carried over cuts the tokens re-read by about 59%. Google and Anthropic both recommend new chats for new topics.

#How much does prompt caching save?

Cached input costs 90 to 97.5% less than normal input at OpenAI, Anthropic and Google. Across a whole workload the saving is smaller because only part of each request repeats; an independent 2026 study of agent workloads measured 41 to 80% lower costs.

#Is ChatGPT Plus cheaper than the API?

It depends on what you use. At API prices, $20 buys about 333 typical exchanges with GPT-6 Astra but over 33,000 with GPT-6 Luna. Heavy users of top models get more from the $20 plan; light users of small models pay far less per use.

#Why do I hit Claude's usage limit so fast?

Claude's limits count how much work each message takes, not just how many you send. Long conversations, attached files, extended thinking and connected tools all use more. Start new chats for new topics, keep files in Projects, and turn off features you don't need.

#Are thinking tokens billed?

Yes. OpenAI, Anthropic and Google all bill a model's hidden reasoning as output tokens, at the output price. Lowering the reasoning effort or thinking level for simple tasks is one of the easiest savings.

#How much cheaper is the Batch API?

Fifty percent at all three companies, for requests that can wait up to 24 hours. It can be combined with prompt caching for even larger savings.

#How we calculated this

Prices are standard list prices from OpenAI, Anthropic and Google, checked October 30, 2026. Every saving we computed ourselves is shown with its assumptions; percentages from published studies link to the source. Token counts in the examples are estimates, since each company counts text slightly differently. We re-check prices regularly and update this page when they change.