LLM Cost Calculator

Request size. Count the input tokens of a typical request, including the system prompt, chat history and any retrieved context, plus the tokens in a typical reply. The presets are starting points: a chatbot turn is small, a coding agent step carries a lot of context.

Volume. Multiply by requests per month. If you only know daily traffic, multiply by 30.

Discounts. Prompt caching bills repeated input at a fraction of the normal rate, and batch processing trades latency for a lower price. Both are off by default so the estimate never looks cheaper than your first invoice.

Four common workloads, priced on a small model (Gemini 3.1 Flash-Lite), a mid-range model (Claude Sonnet 5.5) and a flagship (GPT-6 Astra). Each uses the matching preset in the calculator above.

Customer support chatbot: 50,000 requests a month

About 1,700 short exchanges a day costs $28.75 a month on Gemini 3.1 Flash-Lite, $200 on Claude Sonnet 5.5 and $1,000 on GPT-6 Astra. That is a 35× spread for the same traffic, so test whether a small model answers well enough before paying for a larger one.

Internal RAG assistant: 20,000 requests a month

Each question carries retrieved passages, so input dominates. On Claude Sonnet 5.5 the bill is $260 a month, and 62% of it is input. Retrieving fewer, better passages is the most direct way to cut it. Caching helps only with the fixed part of the prompt, such as instructions, because the passages change with every question. The same workload costs $35.00 on Gemini 3.1 Flash-Lite and $1,300 on GPT-6 Astra.

Coding agent: 5,000 steps a month

An agent resends a long, mostly unchanged context on every step, which is what prompt caching is for. Without caching, Claude Sonnet 5.5 costs $400 a month and GPT-6 Astra $2,000. With 80% of input served from the cache, they drop to $184 and $920. Turn on caching in the calculator and set the cached share to match your agent.

Overnight summaries: 100,000 documents a month

Nobody waits on these results, so batch processing is an easy saving. At standard rates the job costs $290 a month on Gemini 3.1 Flash-Lite and $2,200 on Claude Sonnet 5.5. Through the Batch API it comes to $145 and $1,100.

Cost per user

If you charge for your product, divide the monthly bill by active users to see your margin. The chatbot example is 1,000 users sending 50 messages each a month, so every user costs $0.029 on Gemini 3.1 Flash-Lite, $0.20 on Claude Sonnet 5.5 and $1.00 on GPT-6 Astra. Your heaviest users can send many times the average, so check the cost of the top 10% of users as well before setting a price or a usage limit.

The calculator prices input and output tokens at list rates. A few charges sit outside that, so leave some headroom in your budget and compare the estimate with your first invoice.

• Reasoning tokens. Models that think before they answer bill that thinking as output. For reasoning models, set output tokens per request well above the length of the visible reply.
• Cache writes. Some providers charge a premium the first time a prompt prefix is written to the cache. The calculator applies only the cheaper cached read rate.
• Retries. Timeouts, errors you retry and answers your users regenerate are all billed as new requests.
• Tools and storage. Built-in web search, code execution and file storage are usually priced per call or per day, separately from tokens.
• Processing tiers. Priority or fast modes and regional data-residency endpoints cost more on some providers.
• Tax. Prices are in US dollars before tax.

• Route by difficulty. Send simple requests such as classification, extraction and short replies to a small model, and only the hard ones to a flagship. Prices between the two differ by an order of magnitude or more.
• Cap the reply. Set a maximum output length and ask for concise answers. Output is the more expensive side on almost every model.
• Trim the context. Summarize long chat history, retrieve fewer and better passages, and drop tool definitions a request does not need.
• Keep the prefix stable. Put instructions and reference documents at the start of the prompt and leave them unchanged between requests, so they are served from the cache.
• Batch what can wait. Evaluations, backfills and nightly reports rarely need an answer within seconds.
• Set a spend limit. Most provider consoles let you cap spend and send alerts before a runaway loop turns into a large invoice.