What does your LLM workload actually cost per month?
Enter your traffic and token sizes. Every major model gets priced against the same workload and ranked cheapest first — including prompt caching. It runs entirely in your browser: no signup, no backend, nothing leaves this page.
Workload inputs
Or type your own numbers below — everything updates live.
Total API calls billed in a month.
System prompt + context + user message.
Include reasoning tokens — they bill as output.
Share of input tokens served from prompt cache.
Results summary
All models, priced against your workload
—
| Loading pricing data… | |||||||||
How LLM API pricing actually works
Almost every provider bills the same two things: tokens you send and tokens the model generates. The rates are quoted per million tokens, and the two rates are never equal. Output typically costs three to five times what input costs, because input is processed in a single parallel pass over the whole prompt while output is produced one token at a time, each token requiring its own pass through the model. That asymmetry is the single most useful fact for controlling a bill: a 20,000-token prompt with a 200-token answer is usually cheaper than a 2,000-token prompt with a 2,000-token answer, even though the second request moves far fewer tokens overall.
Reasoning models complicate this. When a model thinks before answering, those thinking tokens are billed as output even when they are summarized or hidden from you entirely. A reasoning model that returns a 300-token answer may have generated several thousand output tokens getting there. If you are pricing a reasoning workload, put your realistic total generation into the output field, not the length of the visible answer, or you will underestimate by an order of magnitude.
Prompt caching
When many requests share a long, identical prefix — a system prompt, a tool schema, a document you keep asking about — providers can cache the processed form of that prefix and skip most of the work on later calls. Cache reads are commonly billed around one tenth of the normal input rate. Writing to the cache costs a premium, often about 25 percent above the input rate, so a cached prefix typically breaks even after two or three reuses and is pure savings after that.
The catch is that caching is a prefix match on exact bytes. A timestamp in your system prompt, a request ID, a tool list that reorders itself, or a JSON object serialized with unstable key order will invalidate everything after the change and silently drop your hit rate to zero. Put stable content first and volatile content last, and verify with whatever cache-hit field your provider returns in its usage payload rather than assuming it works. The cached input share field above is where you model this: set it to the fraction of your input tokens you expect to be served from cache. Models that publish no cache-read rate are priced with all input at the standard rate and flagged in the table.
Context windows and why long context costs more
A context window is a capacity limit, not a price tier — you pay for the tokens you actually send, not for the size of the window. Two things still make long context expensive. First, some providers apply a higher per-token rate above a threshold, commonly 200,000 tokens in a single request. Second, and far more significant in practice, a large window changes behaviour: it becomes tempting to paste an entire codebase or knowledge base into every call, and you pay full input price for all of it on every single request. Retrieval that finds the 4,000 tokens that matter beats sending 100,000 tokens and hoping, on both cost and, usually, accuracy.
Batch discounts and the other levers
Most providers offer an asynchronous batch mode at roughly half of standard pricing, with results delivered within a processing window instead of immediately. For nightly classification, backfills, evaluations, or bulk summarization, that is a 50 percent saving for a change of endpoint. The numbers in this calculator are standard synchronous rates, so halve the monthly figure for anything you can move offline. Beyond that, the levers in rough order of payoff are: cache your stable prefix, stop sending context the model does not need, cap output length, route easy requests to a smaller model, and only then consider switching providers.
Estimating your token counts
For ordinary English prose, one token is about four characters, or roughly 0.75 words — so 1,000 words is around 1,300 to 1,400 tokens and a dense A4 page is about 700. Code, JSON, markup, long identifiers and non-Latin scripts all tokenize worse, frequently 1.5x to 2.5x the prose estimate for the same character count. Tokenizers differ between providers, so the same text will not produce identical counts everywhere. These heuristics are good enough for sizing a budget; before you commit to a contract or an architecture, measure real requests with the provider's own token-counting endpoint or usage response.
A note on accuracy. Prices move, and they sometimes move without an announcement. This tool shows a dataset with the date it was last checked, and links every provider to its own pricing page. Enterprise agreements, regional pricing, and cloud-marketplace rates through AWS, Google Cloud or Azure can all differ from public list price. Use this to compare and to size, then verify before you commit.
Frequently asked questions
How many tokens is my text?
For ordinary English prose, one token is roughly four characters, so about 0.75 words per token — 1,000 words lands near 1,300–1,400 tokens. Code, JSON, non-Latin scripts and long identifiers tokenize worse, often 1.5x to 2.5x the prose estimate for the same character count. Every provider uses a different tokenizer, so treat these as planning numbers and confirm against a real token-counting endpoint before you commit to a budget.
Why is output so much more expensive than input?
Input tokens are processed in one parallel forward pass over the whole prompt, while output tokens are generated one at a time, each requiring its own pass. That serial generation is what occupies the accelerator, so output is typically priced three to five times higher than input. The practical consequence is that verbose answers cost far more than long prompts, and asking for concise output is usually the cheapest optimization available.
Does prompt caching actually save money?
Yes, when a large prefix is genuinely stable and reused. Cache reads are commonly billed at about one tenth of the normal input rate, while writing to the cache costs a premium of roughly 25 percent over the input rate. A cached prefix therefore pays for itself after two or three reuses. It only works if the prefix is byte-identical: a timestamp, a request ID or a reordered JSON key anywhere in the cached region invalidates everything after it.
Does a bigger context window cost more per token?
The advertised per-token rate is usually flat, but some providers charge a higher long-context rate once a single request exceeds a threshold such as 200,000 tokens. The larger effect is behavioural: a big context window invites you to stuff more into every request, and you pay the full input price for all of it on every call. Retrieval that sends 4,000 relevant tokens is almost always cheaper than sending 100,000 tokens and hoping.
What about batch processing discounts?
Most major providers offer an asynchronous batch mode at about 50 percent of standard pricing, with results returned within a processing window rather than in real time. If your workload is offline — nightly classification, backfills, evaluation runs, bulk summarization — halving the bill usually requires nothing more than switching endpoints. This calculator shows standard synchronous pricing, so halve the monthly figure for anything you can move to batch.
Are these prices guaranteed to be current?
No. The dataset carries the date it was last verified, every provider name in the table links to that provider's official pricing page, and every row is marked with how its price was checked — read from the provider's own page, or cross-referenced across independent sources where that page was not reachable. LLM prices change often and sometimes without announcement, and regional, enterprise and cloud-marketplace rates can differ from list price. Use this tool to compare options and size a budget, then confirm the exact numbers with the provider before committing.