AI pricing pages quote a rate per million tokens, which doesn't answer the budgeting question: what does one call cost, and what will it cost at your volume? This calculator converts each model's published rates into a per-request and monthly cost. You can enter sizes in tokens, words, lines of code or characters, and model prompt caching and batch processing, which is where large-document workloads save the most.
Figures checked by Muhammad Ahmad against OpenAI API pricing, Claude API pricing and Gemini API pricing. Prices and rules change, so confirm with the official source before relying on them. How we check the math
How it works
Providers bill input tokens (everything you send: instructions, documents, conversation) and output tokens (the reply) at different rates. Cost per request = input tokens × input price + output tokens × output price, with both prices quoted per 1,000,000 tokens.
If you enter words, lines or characters, they're converted to tokens first. English averages about 1.33 tokens per word; Arabic and Urdu take roughly 2 to 3, because tokenizers are trained mostly on English. Code varies with the language and line length. For exact counts, use the provider's token counter and enter tokens.
Prompt caching charges a much lower rate for input the provider has seen recently, such as a long system prompt or a contract you ask many questions about. Enter the share of each request that repeats and the cached price as a share of the normal price. Batch processing trades speed for a discount on requests that can wait, like overnight document runs.
Monthly cost = cost per request × requests per month. Prices change often: the model list here was checked against official pricing pages on the date shown, and the cache and batch rates are yours to set from the provider's current page.
A worked example
Summarising a 10,000-word English document into a 1,000-word answer is about 13,300 input and 1,330 output tokens. On Claude Sonnet 5 ($2 input, $10 output per million), each call costs $0.0399, or $239.40 a month at 200 calls a day (6,000 a month). If half the input is served from cache at 10% of the input price, the cost drops to $0.02793 per call and $167.58 a month.
More examples
The default 10,000 requests (5,000 input and 1,000 output tokens each) cost $10 a month on GPT-6 Luna, $200 on GPT-6 Sol or Claude Sonnet 5, and $400 on Claude Opus 5.5.
With 80% of each prompt cached, GPT-6 Sol falls from $200 to $128, and output becomes most of the bill.
Common mistakes
- Comparing input prices and ignoring output, which usually costs several times more.
- Forgetting reasoning tokens, billed as output.
- Using old prices; GPT-6 Sol and Luna halved OpenAI's mid and low tiers on 22 September 2026.
- Ignoring caching and batch discounts.
Questions people ask
How many tokens is a word?
In English, about 1.33 tokens per word on average, or roughly 4 characters per token. Short common words are often one token; long or rare words split into several. Arabic, Urdu and many other languages use more tokens per word.
Why do output tokens cost more than input tokens?
Input tokens are processed together in one pass, but output tokens are generated one at a time, each needing its own pass through the model. That extra compute is why output is typically priced several times higher than input.
How much does prompt caching save?
It depends on how much of each request repeats and the provider's cache price. Cached input is often charged at a small fraction of the normal rate, so a workload that resends a long document or system prompt can cut its input cost by more than half. Some providers also charge to write to the cache.
Why are Arabic and Urdu more expensive per word?
Most tokenizers learned their vocabulary mainly from English text, so words in other scripts are split into more pieces. The price per token is the same, but each word costs more tokens, often two to three times as many.
When should I use batch processing?
For work that doesn't need an instant answer: summarising a document archive, tagging support tickets overnight, or running evaluations. Batch requests are usually returned within a day at a discount on the standard price.
Guides
- AI API Pricing in 2027: What's Likely to Change, and WhenFrontier model pricing fell 50-80% through 2026. 2027 will keep cutting prices on commodity-tier models while top-tier reasoning models hold or raise their rates.
- Why AI Agents Cost More Than You Expect (and How to Cut the Bill)An agent re-sends a growing pile of context on every step, so doubling the steps roughly triples the cost. Here's the math, and the four levers that bring it down.
- How to Estimate and Reduce LLM API CostsAPI bills come from input and output tokens priced separately. With three numbers per request you can estimate a monthly bill and compare models before writing any code.
- Tokens vs Words: How AI Models Count TextAI models read and bill text in tokens, not words. A useful rule of thumb is 1,000 tokens to 750 English words. Here is where that comes from and when it breaks down.
- How to Price an AI Feature So It Stays ProfitableStart from cost per active user, not cost per token. Here's how to estimate it, including chat history, and turn it into a price with a healthy margin.
- How Much VRAM Do You Need to Run an LLM Locally?GPU memory for a local model comes from three places: the weights, the KV cache and runtime overhead. Here's how to estimate each, and why quantization matters more than anything else.
- Why Gross Margin Matters More Than Revenue in SaaS Unit EconomicsTwo businesses with the same price and churn can have very different customer value. The difference is gross margin, and it shows up in LTV and payback.
- How Much Does an AI Chatbot Cost? GPT vs Claude vs GeminiEvery new message in a chat resends the whole conversation as input. That makes long conversations much more expensive than the sum of their messages.
- API Rate Limits: Converting RPM to RPS and Leaving HeadroomRate limits are quoted per second, minute, hour or day. Converting them to one unit and adding a safety buffer tells you how fast your code can safely send requests.