How are LLM token costs calculated?

Per-request cost = (input tokens ÷ 1,000,000 × input rate per 1M) + (output tokens ÷ 1,000,000 × output rate per 1M). Each model has its own rates. GPT-5.6 Terra, for example, charges $2.00/$12.00 per 1M tokens, so 1,000 input and 500 output tokens cost $0.002 + $0.006 = $0.008. Two more terms join that sum when they apply: reasoning tokens, billed at the output rate, and cache writes, billed at each model’s own published write rate. Google’s cache storage is the exception: it is charged by the hour rather than per request, so it is added to the monthly figure instead. With cached input or batch pricing turned on under Advanced Pricing Modifiers, those rates apply before the monthly projection. A million input tokens on GPT-5.6 Terra cost $2.00, so 20,000 tokens cost $0.04 and one token costs $0.000002.

Which LLM API is cheapest?

It depends on your token volumes and the quality you need. The lowest per-token rates belong to small models: Amazon Nova Micro ($0.035 input / $0.14 output per 1M tokens), Ministral 3 3B ($0.10 / $0.10) and Amazon Nova Lite ($0.06 / $0.24). GPT-6 Luna ($0.10 / $0.50) and GPT-5.6 Luna ($0.20 / $1.20) are the cheapest OpenAI models in this calculator. GPT-5.6 Terra and Claude Sonnet 5.5 sit in the middle; flagships such as GPT-6 Astra and Claude Fable 5.1 cost $10 / $50, and the Pro models reach $30 / $180. A cheaper model that needs longer prompts or more retries can still cost more overall, so run your own workload in the Compare view, which prices every model you pick on the same volumes and marks the lowest cost per request.

How do I estimate monthly LLM API costs?

Enter your average input and output tokens per request, then set your expected requests per month. The calculator multiplies per-request cost by that volume to give a monthly projection. Turn on cached-input or batch pricing under Advanced Pricing Modifiers if your usage qualifies. The Compare view applies the same volumes to every model you select, so the monthly figures differ only by each model’s rates at that token count. Before launch, start from a scenario preset close to your workload, replace its token sizes and volume with your own estimates, and set the Variability Range to see a low and a high cost per request.

What do the scenario presets do?

A scenario preset fills in an example workload, such as support chat or content generation, so you can start from numbers and change them. The volumes are CalcTrail’s own examples, not measured averages.

Can I compare multiple models with the same workload?

Yes. Press Compare models and the Compare view runs the same token workload across up to five models you pick, ranked by estimated cost per request. It prices only what that screen shows: the token volumes and each model’s own rates, plus what you set for a column under Customize pricing: custom rates, batch pricing or cached input. Each provider counts the same text with its own tokenizer, so check the token count for each model.

Are the model prices up to date?

Rates are read from the provider’s own pricing page, and the model details show the date of the last check. Llama 4 is the exception: Meta’s pricing page lists no Llama 4 rate, so the rates here are the median across hosting providers, and the model details say so. A model the provider has shut down is no longer offered as a choice. It appears only when a saved tab already used it, listed under Deprecated, so the models you can pick are ones the provider still serves. All rates are in USD, the currency these APIs bill in.

What is the difference between input and output tokens?

Input tokens are the text you send (prompts, context, instructions). Output tokens are what the model generates. Output usually costs more: across the current models in this calculator, the output rate runs from the same as the input rate to about 8 times it; the middle of the models listed here is 5 times. OpenAI’s rule of thumb puts a token at about 4 characters or 0.75 words of English.

Why did my cost jump when I increased the input tokens?

You most likely crossed the model's long-context threshold. Some models charge a higher rate for large prompts. GPT-5.6 Terra, for instance, moves from $2.00/$12.00 to $4.00/$18.00 per 1M once input passes 272,000 tokens; the Grok models, Gemini 2.5 Pro and Gemini 3.1 Pro Preview switch at 200,000, and MiniMax M3 at 512,000. The catch is that the higher rate applies to the whole request, not just the tokens past the line. So a prompt one token over the threshold costs up to twice as much as one just under it, depending on how much of the bill is input. The calculator applies this automatically and says so in the adjustments list under your result.

Does prompt caching always save money?

No. Reading from the cache is cheap, 10% of the input rate on most Claude models, but writing to it costs more than a plain input token: 125% of the input rate for a 5-minute cache on Claude, GPT-5.6 and GPT-6 models. A prefix you write and never read back therefore costs 25% more than not caching at all. Kimi K3 is the exception: its 5-minute write costs the plain input rate, so caching there never costs more. Anthropic puts the break-even at one cached read for a 5-minute cache and two for a one-hour cache. The calculator prices the 5-minute write and works its break-even out from each model’s own rates. Google is different again: it bills storage by the hour for as long as the cache is alive, so a cache held overnight costs something even if nothing reads it.

Why is my bill higher than the tokens I can see?

Reasoning tokens can explain it. Models that think before answering generate a private chain of thought that never appears in the response, and OpenAI, Anthropic and Google all bill it at the output rate, the dearer side on most models. A 500-token answer with 4,000 reasoning tokens behind it is billed as 4,500 output tokens. The API reports the real figure per request; put it in the Reasoning Tokens field and the estimate counts it. No provider publishes typical counts, which is why this calculator asks rather than guesses.

Can I use word counts instead of tokens?

Yes. Switch a side to Words and the calculator converts at an estimated ratio for the model you picked: about 1.33 tokens a word for GPT and for models without a ratio of their own, which is OpenAI’s rule of thumb of 0.75 words a token. Claude and Gemini models use their own estimates, listed under Formulas and sources, and “How it works” states the ratio used.

LLM Token Cost Calculator - GPT, Claude & Gemini Pricing

Formulas and sources

Each formula this calculator uses, where it comes from, and what it leaves out. "How it works" shows the same formula with your own numbers.

Cost of one request

In the one-model view

Cost / request = (Input Tokens ÷ 1,000,000 × Input Rate) + (Output Tokens ÷ 1,000,000 × Output Rate)

Cost / request: the cost of one API call
Input Tokens: the prompt, system instructions and context sent to the model
Output Tokens: the tokens the model writes back
Input Rate: the price per 1 million input tokens, after any option below
Output Rate: the price per 1 million output tokens, after any option below

Rates are in US dollars per 1 million tokens, as each provider’s own pricing page states them. For Llama 4, which Meta’s pricing page does not list, they are the median of its hosting providers’ prices. This panel carries no prices: the model details under the model picker show the rate in use, the date it was last checked and where it was read.

Source: Standard formula; no single publisher. Rates are each provider’s own per-1M-token prices, such as OpenAI’s [1], Anthropic’s [2] and Google’s [4].

Monthly cost

With Requests per month set

Monthly = Cost / request × Requests / month

Monthly: the cost of a month of requests
Cost / request: from the formula above
Requests / month: how many requests you send in a month

A Google cache kept by the hour adds its storage charge once to the month, not to each request (see "Cache storage").

Source: Standard formula; no single publisher.

Comparing models

In the Compare view

Monthly = ((Input Tokens ÷ 1,000,000 × that model’s input rate per 1M) + (Output Tokens ÷ 1,000,000 × its output rate per 1M)) × Requests / month

Monthly: each model’s cost for the month
Input Tokens: the same for every model
Output Tokens: the same for every model

Each model is priced at its own rates, and its own long-context tier where one applies. The batch, cache, reasoning and custom-rate settings of the one-model view do not reach the columns; a column’s own Customize pricing applies to that column only. A Words setting does not reach the Compare view either: the numbers there are read as tokens. The lowest cost per request is marked Best.

Source: Standard formula; no single publisher.

Cached input and cache writes

With Cached input pricing on

Input cost = (Input Tokens − Cached Tokens − Written Tokens) ÷ 1,000,000 × Input Rate + Cached Tokens ÷ 1,000,000 × Cached Rate + Written Tokens ÷ 1,000,000 × Write Rate

Input Tokens: all the input tokens of the request
Cached Tokens: input tokens read back from the provider’s cache
Written Tokens: input tokens stored in a 5-minute cache on this request
Input Rate: the model’s input rate
Cached Rate: the model’s published cached price, a fraction of its input rate
Write Rate: the model’s published write price, a fraction of its input rate

A model with no published cached price bills cached tokens at the full input rate, and one with no published write price bills written tokens as plain input. A write only pays off once it is read back: the result says after how many reads. The written tokens count in the Input Cost card.

Source: Standard arithmetic on each provider’s published shares: OpenAI’s Cached input and Cache writes columns [1]; Anthropic’s prompt caching table and its break-even of one read for a 5-minute write [2].

Batch pricing

With Batch pricing on

Batch Rate = Standard Rate × Batch Share

Batch Rate: the rate for input, cached input, cache writes, reasoning and output tokens alike
Standard Rate: the rate without Batch pricing
Batch Share: the model’s published batch price over its standard price

Where a provider publishes a batch price, Batch Share is that price over the standard price. OpenAI, Anthropic and Google set it at half. A model with no published batch price, or no batch API, keeps its standard rates.

Source: Each provider’s own batch prices: OpenAI’s Batch table, which halves the cache-write price too [1], Anthropic’s, whose cache multipliers stack with the Batch discount [2][3], and Google’s [4].

Reasoning tokens

For OpenAI, Anthropic and Google models

Billed Output Tokens = Output Tokens + Reasoning Tokens

Billed Output Tokens: the output tokens the provider charges for
Output Tokens: the visible answer
Reasoning Tokens: the private tokens a model spends thinking before it answers; the figure your API reports

They are not in the response, so the bill is larger than the visible answer.

Source: Billed as output tokens by OpenAI [5], Anthropic [6] and Google [4].

Long-context rates

When Input Tokens pass a model’s threshold

Above a model’s threshold, OpenAI, Google, xAI and MiniMax bill the whole request at their long-context rates, not only the tokens past it. A request just over the line can cost up to twice as much as one just under it. Claude 4.6 and later models cost the same at any length; the one exception Anthropic lists, Claude Haiku 5.5, is not in this calculator yet. The model details show each threshold.

Source: As each provider publishes it: OpenAI’s Long context prices [1], Google’s prompts over 200k tokens [4], and Anthropic’s standard pricing across the full context window [2].

Cache storage

For Google models with Cached input pricing on

Monthly = Cost / request × Requests / month + Cached Tokens ÷ 1,000,000 × Storage Rate × Hours

Monthly: the cost of a month of requests
Cost / request: the cost of one request
Requests / month: requests in the month
Cached Tokens: the tokens kept in the cache
Storage Rate: Google’s storage price per 1 million tokens per hour, Standard tier
Hours: how long the cache is kept in the month

Google charges by the hour for keeping a cache, on top of the cheaper read. It is one charge for the period, not one per request.

Source: Google’s storage price, from its context caching rows [4].

Words to tokens

With a side switched to Words

Tokens = Words × Tokens per Word

Tokens: the estimated token count, rounded up
Words: the words you enter
Tokens per Word: an estimate per model family: about 1.33 for GPT and models without a ratio of their own, 1.52 for Claude 4.6 and earlier, 1.97 for Claude 4.7 and later, 1.67 for Gemini

These are estimates; real counts depend on the text and the provider’s tokenizer. "How it works" states the ratio it used for your model.

Source: Based on OpenAI’s rule of thumb of 0.75 words a token [7] and Anthropic’s 3.5 characters a token [8]. Claude 4.7 and later is scaled by the 30% more tokens Anthropic gives for its newer tokenizer [2]. Gemini’s is CalcTrail’s own estimate, at the high-token end of Google’s 60 to 80 words per 100 tokens [9].

Cost range

With Variability set above 0

Range = Cost / request × (1 ± Variability ÷ 100)

Cost / request: the estimate
Variability: how far you expect the token counts to differ, as a percentage; 20 to start

The range is for one request; the monthly figure carries none.

Source: CalcTrail’s own range; no publisher.

Sources

OpenAI. Pricing, the Standard and Batch tables (Cached input and Cache writes columns) and Long context. Undated page, read October 5, 2026.
Anthropic. Pricing, sections "Prompt caching", "Batch processing" and "Additional notes". Undated page, read October 5, 2026.
Anthropic. Prompt caching, its pricing section and FAQ. Undated page, read October 5, 2026.
Google. Gemini Developer API pricing, the per-model Standard tables: Batch, context caching and its storage price, thinking tokens, prompts over 200k tokens. Undated page, read October 5, 2026.
OpenAI. Reasoning models, section "How reasoning works". Undated page, read October 5, 2026.
Anthropic. Adaptive thinking, section "Pricing". Undated page, read October 5, 2026.
OpenAI. Key concepts, section "Tokens". Undated page, read October 5, 2026.
Anthropic. Glossary, entry "Tokens". Undated page, read October 5, 2026.
Google. Understand and count tokens, its opening section on tokens. Undated page, read October 5, 2026.

What this calculator does not cover

Prices are standard-tier list prices in US dollars, before tax. Service tiers and surcharges are not applied: OpenAI’s Fast and Flex modes and its regional-processing uplift, Google’s Priority tier, and the US-only inference surcharges of Anthropic and xAI.
Only the 5-minute cache write is priced (Anthropic’s 1-hour write costs twice the input rate). Inside a batch, Anthropic serves cache hits on a best-effort basis.
Fees beyond tokens are left out: web search, code execution, the extra input tokens tool use adds, and earlier thinking blocks that Anthropic bills as input.
Time-of-day prices, such as DeepSeek’s off-peak half price, and volume or negotiated discounts are not applied. Custom rates take a contract price.
The Reasoning field is offered for OpenAI, Anthropic and Google models only, the three that say they bill reasoning as output. Other providers’ thinking tokens are not counted.
A price change a provider has already announced takes effect on its date, by your device’s calendar. A model a provider has shut down is not offered for new estimates, and a saved tab keeps the model it used.

Sources checked on October 5, 2026.

Help

Select a Model: Choose the LLM you plan to use (GPT-5.6, Claude, Gemini and others).
Choose Input Mode: Enter tokens directly, or switch a side to Words and the calculator converts them at an estimated ratio.
Enter Input Tokens: Input the number of tokens (or words) for your prompts.
Enter Output Tokens: Input the expected number of tokens (or words) in the response.
Set Requests per Month: Enter the number of requests per month to see a monthly projection. Leave it blank for the cost of one request.
Add Pricing Modifiers (Optional): Open Advanced Pricing Modifiers to turn on Batch pricing or Cached input pricing, and to enter cached, cache write and reasoning tokens where the model supports them. Select Customize pricing to enter your own rates per 1M tokens.
View Cost Breakdown: See the cost per request, the input and output cost, and the effective rate per 1M tokens.
Open the Compare View: Press Compare models. The input tokens, output tokens and requests per month you entered apply to every model. Compare reads them as tokens, so convert any word counts first.
Add Models to Compare: Add the models you are weighing with “Add model”, up to five. Search by name or provider; models no longer offered are hidden.
Read the Comparison: The cheapest model for your workload is highlighted in green with a “Best” badge, and each pricier model shows how much more it costs. Cost / month shows a dash until requests per month is entered.
Share the Comparison: Use Copy or Export to share it, or open “How it works” to see how each cost is worked out.

Understanding LLM Token Costs

What is a token?

LLMs bill by the token, not the word. A token is a chunk of text. OpenAI's rule of thumb is about 4 characters, or three-quarters of a word, of English per token; other providers count differently. Before any of the options below, a request costs its input text plus the output the model generates, each at the model's per-token rate.

Input vs output pricing

Providers charge a separate rate for input (what you send: the prompt, system message, and any context) and output (what the model writes back). Output costs more per token than input on most models listed here, so a long answer can cost more than a long prompt.

Cached input and batch pricing

If you resend the same context, such as a long system prompt or a document, cached-input pricing bills those repeated tokens at the model's published cached rate. That is 10% of the input rate on most Claude models, for example, but no discount at all on GPT-5.5 Pro. Writing to the cache costs extra: Claude, GPT-5.6 and GPT-6 models charge 125% of the input rate for a 5-minute cache write. So a prefix you never read back costs more than not caching at all. Kimi K3 is the exception, charging its plain input rate for a 5-minute write. Google works differently again: it charges to keep a cache alive, by the hour, for as long as you hold it. Batch APIs bill jobs that can wait at a lower published rate, and some models have no batch API. The Advanced Pricing Modifiers panel applies each model's own rates, and leaves a model at its standard price where the provider publishes no cached or batch price.

Reasoning tokens you never see

Reasoning models think before they answer. OpenAI, Anthropic and Google bill those private tokens at the output rate, and they do not appear in the response. A short answer with a long chain of thought can cost many times what its visible length suggests. No provider publishes typical counts, so the Reasoning Tokens field is yours to fill from what the API reports back.

How to reduce token costs

OpenAI's cost optimization guide lists these levers: fewer requests, fewer tokens, a smaller model, and the Batch API or flex processing for work that can wait. Caching repeated context is covered above.

Common mistakes

Counting only the prompt and leaving out output tokens, which cost more per token on most models. Leaving out the system prompt and conversation history, which are billed on every request. Reading a rate per 1K when the provider quotes it per 1M, or the reverse.

LLM Token Cost Calculator - GPT, Claude & Gemini Pricing

Formulas and sources

Cost of one request

Monthly cost

Comparing models

Cached input and cache writes

Batch pricing

Reasoning tokens

Long-context rates

Cache storage

Words to tokens

Cost range

Sources

What this calculator does not cover

Help

Understanding LLM Token Costs

What is a token?

Input vs output pricing

Cached input and batch pricing

Reasoning tokens you never see

How to reduce token costs

Common mistakes

Frequently Asked Questions

Related Calculators

AI ROI Calculator

Image Gen Cost

Speech Cost

Token Converter

LLM Token Cost Calculator - GPT, Claude & Gemini Pricing

Formulas and sources

Cost of one request

Monthly cost

Comparing models

Cached input and cache writes

Batch pricing

Reasoning tokens

Long-context rates

Cache storage

Words to tokens

Cost range

Sources

What this calculator does not cover

Help

Quick Start Guide

Understanding LLM Token Costs

What is a token?

Input vs output pricing

Cached input and batch pricing

Reasoning tokens you never see

How to reduce token costs

Common mistakes

Features

Common Use Cases

Frequently Asked Questions

Related Calculators

AI ROI Calculator

Image Gen Cost

Speech Cost

Token Converter