How much a chatbot costs per month: three traffic levels, eight models

A chatbot that answers customers through a model API costs a fraction of a cent per conversation, but the bill grows with three things you control: how many conversations it handles, how long each one runs and which model answers. This page puts numbers on chatbot cost per month with one worked example at 1,000, 10,000 and 100,000 conversations, on eight current models, using prices read from each provider's own pricing page on October 28, 2026. The figures are our calculation, and every assumption is listed so you can redo it with your own numbers.

The example conversation

The example is a support assistant on a website. Each conversation has four turns. The system prompt, with the assistant's instructions and a short product summary, is 1,200 tokens; each customer message is 50 tokens and each answer is 200 tokens. If tokens are new to you, start with what a token is.

The model has to read the system prompt and the whole conversation so far on every turn, and the provider bills all of it as input tokens. That is why input grows turn after turn while output stays flat:

Table: tokens per turn in the example conversation

Two thousand tokens is well within the limits of these models, but very long chats do eventually run into the context window.

Monthly cost at list prices

Table: monthly API cost of the example chatbot without cache discounts (our calculation, prices of October 28, 2026, rounded)

Four details explain the order of the rows:

• Claude counts more tokens. Anthropic's pricing page says its tokenizer for Claude 4.7 and later models "produces approximately 30% more tokens for the same text". We apply that 30% to both Claude rows, which is why Claude Haiku 5.5 costs more than GPT-6 Luna at the same list price.
• DeepSeek has peak hours. The figures use its peak price. Outside its peak hours (01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday) it charges half.
• One price is about to change. Google lists Gemini 3.8 Flash at $0.75 and $3.75 through December 31, 2026, and at $1.50 and $7.50 from January 1, 2027, which doubles that row.
• Output weighs more than its share. Answers are 11% of the tokens in this example (800 of 7,300) but about 38% of the bill on most rows, because output tokens cost four to six times as much as input tokens on these models.

The spread is wide: the last row costs 26 times the first for exactly the same conversations.

How much prompt caching saves

The system prompt and the earlier turns repeat on every call, which is the case prompt caching was built for: the provider bills the repeated beginning of a prompt at a reduced rate. Our guide to prompt caching explains the mechanism. The rules that matter for this chatbot, from each provider's documentation on October 28, 2026:

• OpenAI caches by default on GPT-5.6 and later, from 1,024 tokens. A cache write costs 1.25 times the normal input price for the default 30-minute lifetime, and a read costs 0.1 times (0.05 on GPT-6.1 Sol and GPT-6.1 Terra).
• Anthropic caches when the request includes cache_control, and its automatic mode moves the cache point forward as a conversation grows. The minimum is 512 tokens on Claude Haiku 5.5 and Claude Sonnet 5.5; a write to the 5-minute cache costs 1.25 times the input price and a read 0.1 times (0.05 on Sonnet 5.5).
• DeepSeek caches by default and stores a cache unit at the end of each user input and of each model answer, so the next turn can reuse everything except the new message. It lists no separate write charge.
• Google lists a minimum of 4,096 tokens for implicit caching on Gemini 3.8 Flash and does not list Gemini 3.1 Flash-Lite in that table. No prompt in this example reaches 4,096 tokens, so both Gemini models are left out below.

Table: best-case monthly cost with prompt caching, cache writes included (our calculation, rounded)

Caching cuts the bill roughly in half, not more, because it only discounts input and output is untouched. DeepSeek saves more because its cache also covers the previous answer and it charges nothing to write. Writes are not free elsewhere: at 100,000 conversations they add $40 to GPT-6.1 Sol and $52 to Claude Sonnet 5.5, about 4% of each total.

"Best case" means the system prompt is already cached from a recent conversation and every turn arrives before the cache expires. At 1,000 conversations a month, one starts about every 43 minutes on average, so the system prompt will often have expired by the next conversation (Anthropic's default cache lasts 5 minutes, OpenAI's at least 30), and the first turn pays a write instead of a read. OpenAI also rounds cached tokens down to a multiple of 128, and DeepSeek describes its cache as best effort. Real savings will be somewhat lower than the table.

What moves the bill the most

Running the same calculation with different inputs shows where to look first:

1. Conversation length. Going from four to eight turns multiplies the cost per conversation by about 2.4 (from $0.0105 to $0.025 on GPT-6.1 Terra), because every turn resends the whole history. Resending only the last few turns plus a short summary keeps long chats under control.
2. Answer length. Answers of 400 tokens instead of 200 raise the cost per conversation by 50% on GPT-6.1 Terra, Claude Haiku 5.5 and Gemini 3.8 Flash, and by 45% on DeepSeek V4.1 Flash. An output limit and an instruction to answer briefly help.
3. Model choice. The factor of 26 between the first and last rows dwarfs everything else. Test whether a small model answers your real questions well enough before paying for a large one.
4. Looked-up content. If the bot retrieves help articles before answering, as in retrieval-augmented generation, those passages are extra input on every turn that uses them.

What this estimate leaves out

The tables cover model API fees only. A working chatbot also needs the software that shows the chat window and stores conversations, hosting, monitoring and someone to maintain it. Other costs scale with use:

• Tools and search. Tool definitions are part of the input on every call that offers them (see function calling), and searching your own documents usually needs embeddings.
• Rate limits. 100,000 conversations a month is about 400,000 API calls, around 13,000 a day, and traffic is rarely even across the day. Check your account's limits in advance with our guide to API rate limits.
• Batch discounts do not apply. A batch API halves prices for jobs that can wait hours, which a live chat cannot.

How to estimate your own bill

1. Measure real conversations. OpenAI, Anthropic and DeepSeek report input, output and cached tokens in the usage data of each response. A week of logs gives you averages per conversation.
2. Multiply by your volume or enter the averages and your cache share in our LLM API cost calculator, which uses the same prices as this page.
3. Count tokens with each provider's own tools before comparing models, because the same text becomes a different number of tokens on each one.
4. Recheck prices monthly. Google has already announced a higher price for Gemini 3.8 Flash from January 1, 2027, and the vendor pages for Claude, OpenAI, Gemini and DeepSeek list every rate.

How we checked

Prices were read on October 28, 2026 from OpenAI's API pricing page, Anthropic's Claude pricing documentation, Google's Gemini Developer API pricing page and DeepSeek's models and pricing page. The caching rules come from each provider's caching documentation, read the same day. The first table uses the same formula as our cost calculator, including Anthropic's 30% token adjustment; the second adds cache writes at 1.25 times the input price for OpenAI and Anthropic and follows DeepSeek's published cache units. We did not test answer quality, speed or rate limits, and real traffic will not match the example exactly.