LLM API cost calculator

A worked example: a support chatbot

Say a chatbot handles 200,000 conversations a month. Each sends about 1,500 tokens in, including a shared system prompt, and gets about 300 tokens back. Half the input is served from cache. Here is the monthly bill across six models, at the prices checked on October 20, 2026.

The cheapest option here, Claude Haiku 5.5, costs $46.50 a month. The most expensive, Claude Fable 5.1, costs $4,538, about 98 times as much for the same traffic. Whether the extra capability is worth that is a quality question the calculator can’t answer, but it does show what you’re paying for the answer.

What moves the bill most

• Output tokens. In this example the response is 17% of the tokens but 66% of the cost on Claude Sonnet 5.5, because output is priced at $10 per million tokens against $2 for input. Shorter answers are the cheapest savings.
• Caching. Serving half the input from cache cuts Claude Sonnet 5.5 from $1,200 to $915 a month, a saving of 24%. It pays off when many requests share a long prefix.
• Batch pricing. For work that can wait, such as summarizing 20,000 documents of about 12,000 tokens each, batch pricing takes Claude Sonnet 5.5 from $640 to $320 a month. It doesn’t suit a chatbot, where people are waiting.
• The model itself. Moving between tiers is the largest lever of all, which is why it pays to test whether a smaller model is good enough for the job.

How the estimate works

How is LLM API cost calculated?

Providers charge per million tokens, with separate rates for input and output. Cost per request is input tokens times the input rate, plus output tokens times the output rate. Multiply by your monthly request volume for a monthly estimate.

How much does prompt caching save?

Cached input is billed at a fraction of the normal input rate, often a tenth or less. It pays off when many requests share a long, repeated prefix such as a system prompt, a document, or a codebase.

What is batch pricing?

Anthropic and OpenAI discount batch API requests by 50% in exchange for results within about 24 hours. It suits summarization, classification, and evaluation jobs that don't need an immediate answer.

What does this calculator leave out?

Extra charges for cache writes, long-context price tiers, tool and search fees, reasoning tokens that you can't predict, and taxes. Treat the result as a planning estimate, not a quote.

I’m Courtney Stephenson, an AI strategy and implementation leader, and I built this tool. If you’re planning an AI rollout and want to talk costs, get in touch.