Most small businesses pay less each month for an AI API than they budget for, because the bill runs by the token, not by the call. At list prices as of October 12, 2026, 500 short chatbot conversations cost about $0.13 to $2.60 a month. Five thousand heavier support chats run $2.75 to $55. The model tier decides where you land.
That spread is the real story. The same workload can cost 20 times more on one model than on another, and most online estimates skip the arithmetic. We ran it ourselves from the official price pages of Anthropic, Google, and OpenAI, using three workloads a small shop might actually run.
Tokens, not calls: how AI API pricing works
AI API pricing is metered in tokens, small chunks of text the model reads (input) and writes (output). Providers quote a price per million tokens, and output costs several times more than input. In English, one token is roughly three-quarters of a word, so 100 tokens is about 75 words, according to OpenAI's help center.
That conversion makes estimates concrete. A 450-word answer is roughly 600 tokens. A 300-word customer question? About 400. Tokenizers differ between vendors, so these are ballpark figures, not invoices.
We priced four models that span the cheap end to the mid-tier, which is where most small-business work sits.
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Provider |
|---|---|---|---|
| Claude Haiku 5.5 (prompts up to 100k tokens) | $0.10 | $0.50 | Anthropic |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | |
| GPT-5.4 mini | $0.75 | $4.50 | OpenAI |
| Claude Sonnet 5.5 | $2.00 | $10.00 | Anthropic |
Standard (non-batch) list prices as of October 12, 2026.
The premium tier costs far more. Anthropic's pricing page lists Opus 5.5 at $4/$20 and Fable 5.1 at $10/$50 per million tokens, and OpenAI's gpt-6-astra sits at $10/$50. Our read: few routine business tasks need that tier.
How much does an AI API cost a small business each month?
A small business running a basic AI chatbot or document tool pays anywhere from a few cents to about $300 a month in API fees, based on our calculations from October 12, 2026 list prices. Volume matters, but the model matters more: Claude Sonnet 5.5 costs exactly 20 times what Claude Haiku 5.5 costs on every job below.
The formula is plain. Monthly requests × tokens per request ÷ 1,000,000 × price, done once for input and once for output, then added together. These numbers leave out caching, batching, and retries.
| Workload | Requests/month | Tokens per request (in/out) | Haiku 5.5 | Flash-Lite 3.5 | GPT-5.4 mini | Sonnet 5.5 |
|---|---|---|---|---|---|---|
| A. FAQ chatbot | 500 | 600 / 400 | $0.13 | $0.59 | $1.13 | $2.60 |
| B. Support chat with docs and history | 5,000 | 3,000 / 500 | $2.75 | $10.75 | $22.50 | $55.00 |
| C. Invoice and document extraction | 50,000 | 2,000 / 200 | $15.00 | $55.00 | $120.00 | $300.00 |
Workload B shows the math in full. 5,000 chats × 3,000 input tokens is 15 million tokens in, and 5,000 × 500 is 2.5 million out. On Gemini 3.5 Flash-Lite that's 15 × $0.30 + 2.5 × $2.50 = $10.75. On Sonnet 5.5 it's 15 × $2 + 2.5 × $10 = $55.
Why does B carry 3,000 input tokens when A carries 600? A real support bot sends more than the customer's question. It pastes in reference text from your help docs plus the earlier turns of the conversation, so input piles up while output stays modest.
Per request, the numbers look almost silly. A workload-A chat costs about $0.00026 on Haiku 5.5 and $0.0052 on Sonnet 5.5. Even one document pushed through Sonnet in workload C comes to about $0.006.
Our read: at workload A volume, model price barely matters, since $2.60 is a rounding error. At workload C volume, the tier you pick is a $285-a-month decision.
What pushes the real bill above the estimate
An AI API bill grows when every request carries extra text. System instructions, pasted documents, and conversation history all count as input tokens each time they're sent. Prices also change, sometimes on a published schedule, so a budget built in October 2026 can drift within months.
Conversation history is the quiet one. Each new turn resends the earlier turns, which makes a long chat more expensive per message than a short one. One cost-tracking blog estimates that a 500- to 1,500-token system prompt gets billed on every turn, and that carrying 10 prior messages can push a 500-token request to around 3,000. We could not confirm those figures on any vendor's page, but the mechanism is simply how per-token billing works.
Scheduled price moves matter too. Google's pricing page says Gemini 3.8 Flash costs $0.75 input and $3.75 output per million tokens through December 31, 2026, then $1.50 and $7.50 from January 1, 2027. Same model. Double the bill.
Our figures cover tokens only. Hosting, the platform you build on, and staff time spent checking outputs aren't included. A consumer subscription such as ChatGPT Plus is reported to be billed separately from API usage, so it doesn't offset any of this.
One agency puts typical small-business AI spending at $5 to $40 a month. That roughly lines up with workloads A and B on the cheaper models, though the agency doesn't show its math.
Where the savings come from: batch, caching, and free tiers
The biggest cuts come from three places: batch processing, prompt caching, and free tiers. Anthropic and Google both charge 50% less through their Batch APIs for jobs that don't need an instant answer, and OpenAI's batch prices work out to about half of its standard rates.
Batch fits workload C well. Invoices don't need a reply in two seconds. Running that job through Anthropic's Batch API cuts Sonnet 5.5 from $300 to $150 a month, and Haiku 5.5 from $15 to $7.50. A live chatbot can't wait for a batch, though, so A and B don't qualify.
Caching helps when the same text repeats in every request, like a system prompt or a help-doc excerpt. Anthropic lists cache hits on Sonnet 5.5 at $0.10 per million input tokens versus $2.00 standard. On GPT-5.4 mini, OpenAI charges $0.075 for cached input against $0.75.
We didn't model caching in the table, because the savings depend on how much of each prompt repeats. For workload B, where most input is repeated context, the gap is wide.
Free tiers exist, with limits. Google offers a free tier on its Gemini Flash and Flash-Lite models, with no charge for input or output, while Gemini Pro has none. For a pilot or a very small bot, that can mean a $0 bill.
At a glance
| If your workload looks like… | Monthly cost (Oct. 12, 2026 list prices) | Watch for |
|---|---|---|
| A few hundred FAQ chats | Under $3 on all four models | Not much; model choice barely moves the total |
| A few thousand support chats with docs | About $2.75 to $55 | Conversation history inflating input tokens |
| Tens of thousands of documents | About $15 to $300, half that with batch | The 20x gap between Haiku 5.5 and Sonnet 5.5 |
| Any of the above in 2027 | Recheck vendor price pages | Scheduled hikes, like Gemini 3.8 Flash doubling on January 1, 2027 |
Related on The Daily Sift: Circular AI Deals Explained: Who Pays Whom in 2026 · Why Anthropic's CEO Asked AI Labs to Slow Down
Sources
- Anthropic — Claude API pricing
- Google — Gemini API pricing
- OpenAI — API pricing
- OpenAI Help Center — What are tokens and how to count them
This article is for general information about AI API pricing for small businesses only and is not investment, financial, or tax advice. Figures such as rates, returns, and rules are as of October 12, 2026 and may change. Any financial decision is your own responsibility; please verify with official sources or a licensed professional before acting.