AI Model Prices & API Costs

What Does One AI API Request Cost? Worked Examples

A single AI API request usually costs a fraction of a cent. On our price board on 2026-11-21, a typical customer-support reply (3,000 tokens in, 400 out) cost between $0.0005 and $0.05, depending on the model. That 100-fold gap means the model you choose matters far more than trimming each prompt.

Cover reading "100×. Same job. Wild price gap. Data as of 2026-11-21." beside a paper airplane and a giant gold rocket each carrying the same envelope (AI-generated illustration)
AI-generated illustration

You’ve built a small chatbot and the first invoice is on its way. Will it be pocket change, or more than you planned?

This article prices four everyday tasks on all 42 models we track, then shows how to do the same sum for your own app.

How a Request Is Priced

Every request has two parts, and each has its own price per 1 million tokens:

  • Input tokens: everything you send. That includes your instructions, the conversation so far, any documents and the user’s message.
  • Output tokens: everything the model writes back.

A token is a small piece of text. OpenAI’s help page gives a rule of thumb for English: about four characters, or roughly three quarters of a word. The cost of one request is then:

Cost = (input tokens × input price + output tokens × output price) ÷ 1,000,000

Output almost always costs more than input. On our board the median output price is five times the median input price ($8.75 against $1.75 per 1M tokens). Our guide How LLM API Pricing Works covers the background: tokens, tiers, caching and batch discounts. This article is the hands-on part: real tasks, real numbers.

Four Everyday Tasks

We picked four workloads that cover most small apps. The token counts are round estimates of typical sizes, not measurements of any one product.

Task Input tokens Output tokens What it stands for
A quick question 200 150 A one-line question and a short answer
Support reply with history 3,000 400 A help-desk bot with instructions and a few earlier messages
Summary of a 20-page document 12,000 800 About 9,000 words in, a one-page summary out
Code review of one file 8,000 1,500 A source file plus instructions in, a detailed review out

Which of these four tasks looks most like your app?

What Each Task Costs Across 42 Models

Here are the cheapest, median and most expensive costs for one request at standard list prices, using the prices on our board on 2026-11-21:

Task Cheapest Median of 42 models Most expensive
A quick question $0.00007 (Ministral 3 14B) $0.0016 $0.0095 (GPT-6 Astra)
Support reply with history $0.0005 (GPT-6 Luna, Claude Haiku 5.5) $0.0081 $0.05 (GPT-6 Astra)
Summary of a 20-page document $0.0016 (GPT-6 Luna, Claude Haiku 5.5) $0.026 $0.16 (GPT-6 Astra)
Code review of one file $0.0016 (GPT-6 Luna, Claude Haiku 5.5) $0.025 $0.155 (GPT-6 Astra)

Two things stand out. First, even the most expensive model costs less than 20 cents per request for these tasks. Second, the spread is wide: for three of the four tasks, the most expensive model costs 100 times as much as the cheapest. For the quick question the gap is even bigger, because Ministral 3 14B charges the same low price ($0.20 per 1M tokens) for input and output.

Did you know? Even the most expensive model on our board charges less than 20 cents to read a 20-page document and write a one-page summary. The cheapest charges less than one fifth of a cent for the same job.

A Closer Look at 12 Models

Totals for one model at a time are easier to compare. The table below shows 12 models from different price levels and providers. The last column scales the support reply to 1,000 requests, roughly what a small help-desk bot might handle in a day.

Model Input / output price (per 1M tokens) Quick question Support reply 20-page summary 1,000 support replies
GPT-6 Luna $0.10 / $0.50 $0.0001 $0.0005 $0.0016 $0.50
Claude Haiku 5.5 $0.10 / $0.50 $0.0001 $0.0005 $0.0016 $0.50
Mistral Small 4 $0.15 / $0.60 $0.0001 $0.0007 $0.0023 $0.69
DeepSeek V4.1 Flash $0.30 / $1.20 $0.0002 $0.0014 $0.0046 $1.38
Mistral Large 4 $0.70 / $2.10 $0.0005 $0.0029 $0.0101 $2.94
Gemini 3.8 Flash $0.75 / $3.75 $0.0007 $0.0038 $0.0120 $3.75
Grok 4.7 $2 / $6 $0.0013 $0.0084 $0.0288 $8.40
Claude Sonnet 5.5 $2 / $10 $0.0019 $0.0100 $0.0320 $10.00
GPT-6.2 Sol $2 / $10 $0.0019 $0.0100 $0.0320 $10.00
Gemini 4 Neon $3 / $15 $0.0029 $0.0150 $0.0480 $15.00
Claude Opus 5.5 $4 / $20 $0.0038 $0.0200 $0.0640 $20.00
GPT-6.2 Astra $8 / $40 $0.0076 $0.0400 $0.1280 $40.00

Costs are rounded. Each model’s own page on our board, such as Claude Haiku 5.5, lists its full price tiers and the official pricing page we read them from.

Why Your Real Bill Can Differ

The tables use standard list prices. Your actual cost can be lower or higher than that:

  • Cheaper with caching. If many requests start with the same long text (a system prompt or a policy document), most providers bill that repeated part at a cached-input price. On our board, 38 of 42 models list one, and the median cached price is 90% below normal input.
  • Cheaper with batch processing. Jobs that can wait hours often cost half as much. OpenAI, Anthropic and Google all list batch prices at 50% of standard for the models on our board.
  • Dearer with long prompts. Some models charge a higher rate once a prompt passes a size limit. OpenAI’s current models on our board switch to a long-context price above 272,000 input tokens, and Claude Haiku 5.5 does so above 100,000 tokens.
  • Dearer with reasoning. Reasoning models “think” before they answer. OpenAI’s reasoning guide says those hidden reasoning tokens are billed as output tokens, so a short visible answer can still use many output tokens.
  • Dearer with retries and tools. Failed calls you retry, tool calls and multi-step agents all add requests that a single-request estimate leaves out.

Prices also change. Two examples from the official pages behind our board:

  • OpenAI marks GPT-5.6 Sol’s price as promotional until 2026-11-30.
  • Google lists a price rise for Gemini 4 Carbon and Gemini 4 Argon from 2026-12-01.

Our board keeps the history of every price, with the date we read it.

How to Estimate Your Own Cost per Request

  1. Measure a real request. Most APIs return the input and output token counts in each response. Log them for a day of normal use; don’t guess.
  2. Use typical values, not best cases. Take the median input and output size. If a few requests are very large, price them separately.
  3. Apply the formula above to two or three candidate models on the price board, or use the calculator on that page.
  4. Multiply by volume. Requests per day × 30 gives a monthly figure.
  5. Subtract what qualifies for caching or batch, and add a margin of 10–20% for retries and growth.

As a check: 1,000 support replies a day on a $2 / $10 model costs about $10 a day, or about $300 a month. The same traffic on a $0.10 / $0.50 model costs about $15 a month. If the smaller model’s answers are good enough for your users, that difference pays for a lot of testing. Our collection of cheap models for a chatbot is a good place to start comparing.

What This Means for You

  • Developers: log the real input and output token counts for a day of traffic, then price them on two or three models before you choose.
  • Founders and team leads: the model choice is your biggest cost lever, up to 100 times for the same task, so make it a deliberate decision rather than a default.
  • Curious beginners: one AI answer usually costs less than a cent. Costs only add up when an app sends thousands of requests a day.

Copy-Ready Cost Worksheet

Copy this into a note and fill in the blanks:

  • Typical input tokens per request: ____
  • Typical output tokens per request: ____
  • Requests per day: ____
  • Model A input / output price per 1M tokens: $____ / $____
  • Model B input / output price per 1M tokens: $____ / $____
  • Daily cost = (input × input price + output × output price) ÷ 1,000,000 × requests
  • Monthly cost = daily cost × 30, plus 10–20% for retries and growth

Bottom Line

  • One request costs fractions of a cent on most models. The total grows with volume, prompt length and output length.
  • The choice of model changes the bill by up to 100 times for the same task, more than any prompt trimming will.
  • Measure your real token counts, price two or three models on our board, and test the cheapest one that meets your quality bar first.

All prices in this article are standard list prices in US dollars per 1 million tokens, as shown on our board on 2026-11-21. Our methodology page explains how we read and check them.

Next, our checklist for choosing a model helps you test the cheaper options. Want these updates by email? Sign up here, or follow new guides and data stories through our RSS feed.

Get the Weekly AI News Bay Email

The week’s biggest AI stories, model price changes and outages in one short email. The first issue goes out in early 2027; sign up now to get it.

Double opt-in: we email you a link to confirm. One-click unsubscribe in every email. Privacy policy.

Sources

Written by AI News Bay’s AI writers and checked by a separate AI auditor before publishing. Prices and uptime figures come from our own measurements. Spotted something out of date? Tell us.

Related on AI News Bay

More in AI Model Prices & API Costs

  • How LLM API Pricing Works (Tokens, Tiers, Caching)

    What you pay for when you call an AI model: input and output tokens, long-prompt tiers, cached input and batch discounts, with worked examples.

    3 min read · Nov 21, 2026

Type a name, a topic or a format. Enter opens the full results.