LLM cost calculator: estimate token and inference costs

An LLM cost calculator turns three inputs (model, tokens per request, and monthly volume) into a number your finance team can actually plan around. That number deserves respect, because LLM economics are quietly counterintuitive.

The price of any single model keeps dropping. Sam Altman has noted the cost to use a given level of AI falls roughly 10x every 12 months, and Andreessen Horowitz’s LLMflation analysis puts a hard number on it: GPT-3-quality output fell from $60 per million tokens in 2021 to $0.06 in 2024, a 1,000x drop in three years.

And yet the invoice still climbs every quarter. Epoch AI’s inference-price research puts the fastest bound on it: the cost of a fixed level of LLM performance has fallen anywhere from about 10x to nearly two orders of magnitude a year depending on the task (a median near 50x), which brackets Altman’s 10x estimate at the low end. Bills still go up anyway.

The reason is simple: cheaper tokens get used far more often. Per unit, dirt cheap. In aggregate, a line item with opinions.

So before the feature ships and the meter starts spinning, it pays to do the math. This guide hands you the formula, a working calculator, the current per-token rates from every major provider (each linked to the vendor’s own page), real monthly examples, and the levers that actually lower LLM costs without quietly wrecking quality.

It is the spreadsheet your finance partner wishes someone had built before launch, not after the first surprising bill.

What is an LLM cost calculator?

An LLM cost calculator is a tool that estimates what you will pay to run a language model in production, based on token usage instead of a flat subscription. It is the AI version of a utility-bill estimator. You tell it roughly how much you will use, it tells you roughly what you will owe.

Most of these tools, including this one, double as an AI inference calculator, because inference is where the recurring spend lives. Inference is the model generating a response. Training the model is a one-time cost, and for almost everyone building on an API, that is the provider’s bill, not yours. Inference is the part that repeats on every single request, forever, at the speed of your adoption curve.

You need a calculator for this because LLM pricing refuses to be intuitive. Costs split between input and output tokens, swing more than 100x between models, and bend around caching, batching, and context length. A back-of-napkin guess is usually wrong by a margin your CFO will not find charming.

How do you calculate LLM inference cost?

Multiply input tokens by request volume and the model’s input rate, do the same for output tokens at the output rate, then add the two. That is the whole formula, and it is why a token cost calculator beats intuition every time:

monthly cost = (input tokens x requests x input rate) + (output tokens x requests x output rate)

Rates are quoted per million tokens (MTok). One token is about four characters of English, so a million tokens is roughly 750,000 words, or ten paperbacks. Both your prompt and the model’s reply are metered, which is exactly why a chatty model with a bloated system prompt can cost double its sticker rate.

The classic miss is pricing the visible question and forgetting the hidden input. System prompts, retrieved documents, conversation history, and tool definitions all count as input on every call.

The model answers one question after silently reading ten pages of context you are paying for. A good LLM token calculator counts those pages, not just the question.

Estimate your LLM spend (the calculator)

Pick your model, enter average input and output tokens per request, add your monthly request volume, and the calculator returns a live monthly estimate. Drag the cache-hit slider to watch reuse cut the bill, because for most production apps that one slider moves the number more than anything else.

To ground the tool, here is one workload (a customer-facing assistant at 1,000,000 requests a month, 1,500 input and 500 output tokens each) priced across three Claude tiers at Anthropic’s current standard rates:

Same code, same prompts, a 5x swing in spend, decided purely by which model you select.

Now cache that reused system prompt at a 75 percent hit rate and Sonnet’s bill falls from about $12,000 to roughly $9,000 a month, with nothing changing for the user. Model choice and caching are the two real dials. Everything else is rounding error wearing a costume.

How much does each major LLM cost per million tokens?

Per-token rates span more than 100x between the cheapest budget model and the priciest flagship, so your estimate depends almost entirely on which row you live in.

This table is the LLM cost comparison and the working LLM pricing calculator reference:

This is LLM API pricing in one view, and the spread is the whole story.

DeepSeek V4 Flash at $0.14 input is roughly 35x cheaper than GPT-5.5 ($5). Gemini 3.1 Pro at $2/$12 sits just under GPT-5.4’s $2.50/$15. The distance between a budget model and a flagship is wide enough to flip a feature’s margin on its own, which is the entire case for matching the model to the task instead of sending everything to the most expensive option out of habit.

Want the full per-model detail, including cached and batch rates? The dedicated guides cover OpenAI API pricing, Claude API pricing, Cursor pricing, and Azure OpenAI pricing. If you are pricing a chat product rather than the API, how much ChatGPT costs breaks down the consumer side.

Why does output cost more than input?

On every major provider, output tokens cost about five to six times more than input tokens.

GPT-5.5 charges $5 in and $30 out, a 6x ratio. Claude holds a tidy 5x across its lineup. Generating text simply burns more compute than reading it does.

The practical takeaway lives in your prompt design. A read-heavy job (summarizing documents, answering questions about a codebase) is far cheaper than a generation-heavy one (writing long reports), even at the same total token count. When a cost per token estimate looks scary, the output column is almost always the villain, and a max-tokens cap is the cheapest fix you will ever ship.

What makes your LLM cost estimate jump?

Most budget surprises are not price changes, they are volume changes nobody plugged into the calculator. Four factors move your estimate the most, and they usually stack on top of each other like a very expensive club sandwich.

Context length is first. Both OpenAI and Google charge premium rates past a threshold. GPT-5.5 doubles its input rate above 272K tokens. Gemini 3.1 Pro steps from $2 to $4 input above 200K. Drop an entire codebase into one prompt and you pay double for the pleasure.

Retrieval is second. RAG sharpens answers by stuffing documents into the prompt, which is great for accuracy and brutal for token counts, since every retrieved chunk is billable input on every query. Retries and agents are third and fourth. One coding-agent task can push 400,000 to 2,000,000 cumulative input tokens, and a retry storm turns a single failure into a budget event. The architecture writes the bill, not the user.

How do you lower LLM costs?

The two dials this calculator gives you, model choice and cache-hit rate, are also the two biggest levers to lower LLM costs, often shifting effective spend by half or more. They are not exotic. Most teams just never touch them, then act surprised at the invoice.

Model routing is the first dial. Send classification, extraction, and routing to a cheap tier (Haiku 4.5, GPT-5.4-nano, or DeepSeek V4 Flash) and reserve the flagship for genuinely hard reasoning. As a rule of thumb, routing simple traffic away from the flagship tends to take well over half off the bill.

Caching is the second, and usually the heavyweight. Reused prompt content (system instructions, examples, document context) bills at roughly 10 percent of the standard input rate once cached. On DeepSeek V4 Flash, a cache hit drops input from $0.14 to $0.0028, a 98% cut that is almost rude. Batching adds a flat 50 percent off asynchronous work at most providers.

That is the estimate-side view. For the full operational playbook (semantic caching, quantization, context trimming, concurrency controls, and the rest), the deep dive on inference cost runs through eight strategies, so this page can stay focused on the number.

What does a real LLM workload cost per month?

Quick answer: A typical support assistant runs about $4,000 a month on a budget model and $20,000 on a flagship, same traffic, different row in the table. Abstractions do not pay invoices, so here are three grounded scenarios.

A high-volume support bot (1M requests a month, 1,500 in and 500 out) lands near $4,000 on Haiku 4.5 and $12,000 on Sonnet 4.6, before caching shaves about 25 percent off the input side.

A document pipeline that is input-heavy (10M input tokens a day) is exactly where cheap models and batching earn their keep, often cutting the bill by an order of magnitude.

A coding agent is the wild card. The median looks tame, but heavy automation can hit $500 to $2,000 per engineer a month, because a coding agent bills by the token, not the seat.

The pattern across all three is the same. The median request is cheap and the tail is where the money hides. Model the 95th-percentile user, not the average one, or the launch you are celebrating becomes the invoice you are explaining.

How do you forecast LLM cost before you launch?

Forecast it from four inputs: average tokens per request, expected request volume, your model mix, and a realistic cache-hit rate, then pressure-test the total against your worst-case adoption curve, not your average one. The goal is a number you can defend and a ceiling you can enforce.

Build the base case on expected per-request tokens and volume. Then run a second case at three to five times the volume with a lower cache-hit rate, because a good launch tends to meet the 95th-percentile user, not the median one finance approved the budget for. Two numbers, one optimistic and one honest, beat a single hopeful guess.

Three inputs deserve extra suspicion, since they are the ones teams routinely underestimate. Output length, because output tokens cost several times input and uncapped responses balloon fast. Retrieval depth, because every chunk a RAG step injects is billable input on every query. And call fan-out, because agentic features turn one user action into many model calls, so the forecast has to count calls, not clicks.

Then bolt a ceiling to the estimate. A budget without enforcement is a wish, so pair the forecast with concurrency limits, per-feature spend alerts, and a kill-switch threshold. Teams that forecast AI spend well do not predict the future perfectly. They cap the downside, instrument the rest, and keep the number inside the range they showed finance.

How do you turn LLM spend into AI ROI?

Cutting the bill is a one-time win. Knowing what each customer, feature, and request costs you is a permanent one, and it is the difference between absorbing AI spend and proving AI ROI. A lower number you cannot attribute is still a number you cannot manage.

This is where most teams stall, and where finance has the most to gain. CloudZero’s 2026 AI ROI survey of 260 finance leaders found that 60% admit they are spending more on AI than they can justify with measurable outcomes, while only 16% can see their AI spend in real time. The gap that hurts is not the token rate. It is visibility.

The companies winning with AI are not the ones with the lowest token rate. They are the ones who can say which feature, for which customer, has a 70 percent gross margin and which has 12, and why.

CloudZero is the only AI ROI company that accurately maps token-level usage and cloud cost to the dimensions finance actually cares about: cost per customer, per feature, and per model, across providers rather than one invoice at a time.

A calculator tells you what a workload should cost. This tells you what it does cost, and turns a scary line item into unit economics you can price and defend.

Get a free demo today to see how leading global organizations such as Toyota, Duolingo, Coinbase, Grammarly and more use CloudZero to maximize their AI ROI.

Frequently asked questions

Author Spotlight

Lyne Carolyne

Lyne Carolyne has several years of experience in AI and cloud economics and brings that understanding into the content she creates. Outside work, she's an avid explorer.