How Much Do AI Tokens Cost? A Plain-English Guide for Small Businesses

The question how much do AI tokens cost? begins with a price that looks microscopic.

A few dollars for a million tokens. Fractions of a cent for a short request. It is easy to glance at an AI pricing page and decide that the cost is too small to deserve your attention.

Then the workflow starts sending the same brief again. It attaches a long document, searches the web, retries an answer, calls a second model and produces an output you still have to check. The token price was accurate. Your assumption about the completed job was not.

So, how much do AI tokens cost? The direct answer is: from a few cents to several dollars per million input tokens, and often much more for a million output tokens, depending on the model and provider. But a business should not stop at the price per million. It should calculate the cost of one approved, finished task—including repeated context, output, tools, retries and human review.

That is the difference between reading a menu and receiving the bill.

If you want to turn this calculation into a controlled AI workflow, Ben Angel's 28-Day AI Mastery course helps you choose the right model for the job, define the human approval line and measure the result without becoming an AI engineer.

In This Article

• How much AI tokens cost—and what a token actually is

• The Four-Line Token Receipt

• A current, dated pricing example

• Why the cheapest tokens can create a more expensive job

• A 15-minute token-cost audit

• What the Lex case study proves—and what it does not

• The buying decision to make before you build

• AI token cost FAQs

How Much Do AI Tokens Cost? The Direct Answer

The price per million is only the menu; the finished business task is the bill.

A token is a small unit of text that an AI model reads or generates. It is not exactly a word. OpenAI's current token guide says one token is roughly four English characters or about three-quarters of a word, although the result varies by language, symbols and formatting. Google's Gemini token guide makes the same practical point: your prompt, files, conversation history and the model's response all contribute to usage.

Most API invoices separate at least three lines:

• Input tokens: the instructions and context sent to the model.

• Cached input tokens: repeated context the provider can reuse at a discounted rate when the platform and request qualify.

• Output tokens: the answer generated by the model, usually priced higher than input.

Some systems also charge for tools, searches, files, images, audio or a model's hidden reasoning work. An agent may make several model calls to finish what looked like one request.

That is why “$2 per million tokens” is not a complete answer. Which tokens? On which model? How many times is the context sent? How long is the output? What else does the workflow call? How many attempts produce an accepted result?

My doctrine is simple:

Put the accepted outcome on the receipt.

This page owns the unit-economics question: how metered model usage becomes a task cost. The broader AI cost for small business guide owns the full stack—subscriptions, setup, rework and owner supervision. The AI subscription audit owns flat-plan renewals. And the separate guide to controlling AI agent costs owns step, retry and tool guardrails.

The distinction matters because a ChatGPT, Claude or Gemini product subscription is not automatically the same thing as API access. A subscription typically buys an interface and a set of product limits. An API lets software call a model and commonly bills by usage. Check the exact product and account before you assume one payment covers the other.

The Four-Line Token Receipt

Cost the complete accepted task across input, output, tools and every attempt.

Forget the provider's giant comparison table for a moment. Give one business task a four-line receipt:

1. Fresh input: instructions, new data and uncached context sent during the task.

2. Cached input: stable material reused at a qualifying cached rate.

3. Output: all generated text, including discarded attempts.

4. Multipliers: retries, model calls, searches, tools and the number of times the task runs.

The base formula is:

Task cost = (input tokens ÷ 1,000,000 × input rate) + (cached tokens ÷ 1,000,000 × cached rate) + (output tokens ÷ 1,000,000 × output rate) + tool charges.

Then multiply by every model call required to produce one approved result.

Imagine a weekly research brief using 150,000 input tokens and 15,000 output tokens. Those numbers can arrive faster than you think when the workflow includes a long report, brand guidance, previous work and source material. If it runs once, the cost may be modest. If an agent repeats it for 100 clients, tries twice on average and adds web searches, the bill changes shape.

The useful denominator is not “cost per run.” It is cost per approved brief.

If 100 runs produce 75 briefs your team accepts, divide the entire spend—including failed runs—by 75. That is the business unit cost. The AI tool evaluation framework applies the same discipline before a product earns a permanent place in the workflow.

A Current Pricing Example: Three Models, One Business Task

A lower rate matters only when the model still clears the business acceptance standard.

Prices change, so treat this as a dated example, not a permanent rate card. On August 27, 2026, OpenAI's official model comparison listed these standard prices per one million tokens:

• GPT-5.6 Sol: $4 input, $0.40 cached input and $20 output.

• GPT-5.6 Terra: $2 input, $0.20 cached input and $12 output.

• GPT-5.6 Luna: $0.20 input, $0.02 cached input and $1.20 output.

For our 150,000-input, 15,000-output brief, with no caching or tool charges, the estimated model cost would be:

• Sol: $0.60 input + $0.30 output = $0.90 per run.

• Terra: $0.30 input + $0.18 output = $0.48 per run.

• Luna: $0.03 input + $0.018 output = $0.048 per run.

At 100 runs, those model charges become about $90, $48 and $4.80 respectively. The arithmetic is straightforward. The decision is not.

If Luna needs three attempts and extensive correction while Terra produces an accepted brief in one attempt, the cheaper rate may not create the cheaper finished job. If Sol materially improves a consequential analysis, the additional 42 cents over Terra may be trivial beside the cost of a weak decision.

Google's official Gemini API pricing likewise separates input, cached input and output, with rates that vary by model, prompt size, usage tier and sometimes batch processing. Anthropic's Claude pricing documentation separates model, prompt-caching and tool costs. Always date-stamp the rate you use and preserve the provider link beside your calculation.

Do not paste these example prices into a forecast and forget them. Providers change models, discounts, context thresholds and tool charges. Your spreadsheet should contain a “rate checked” date and an owner responsible for updating it before a purchasing or scaling decision.

Why the Cheapest Tokens Can Create a More Expensive Job

Workflow design decides how many tokens you consume and whether the result is useful.

Token price is an ingredient. Workflow design determines how much of it you consume and whether the result is useful.

Five multipliers commonly turn a cheap request into an expensive system.

1. Repeated context. A long system prompt, document or conversation history may be sent again on every call. Caching can reduce eligible repeated-input cost, but only if the workflow actually uses it correctly.

2. Long output. Output is often more expensive than input. Asking for exhaustive explanations, multiple variants or verbose internal notes can cost more than the source material.

3. Retries and loops. A failed structured output, a weak instruction or an agent without a finish condition can repeat the job. A per-call cap does not protect the complete run.

4. Premium-model defaulting. Routing every classification, extraction and rewrite to the strongest model wastes money. Reserve expensive reasoning for the decisions where it changes acceptance, risk or revenue.

5. Human correction. A six-cent draft that consumes forty minutes of founder review is not a six-cent result. The model invoice and the operating cost must remain separate—and both must be visible.

This is why AI workflows need a completion standard and why an AI agent versus automation decision matters. A fixed automation can be easier to predict. An agent has more freedom to decide which steps and tools to use, so it needs a budget, retry ceiling and stop rule around the whole job.

There is also a quality multiplier most pricing calculators ignore. If a cheaper model generates plausible errors that survive review, the cost appears later as rework, customer confusion or a bad decision. The goal is not minimum token spend. It is the lowest defensible cost for an accepted outcome.

Run a 15-Minute Token-Cost Audit

Change one condition and compare acceptance, total cost and review time.

Choose one repeated AI task. Do not audit the entire company at once.

Minutes 0–3: Name the finished job.

Write the result in observable language: “one source-verified newsletter approved for scheduling,” not “use AI for marketing.” Name the reviewer and the acceptance standard.

Minutes 3–6: Open real usage evidence.

Use the provider's usage dashboard, API response metadata or billing export. OpenAI's usage dashboard guidance explains where organizations can inspect usage and cost. Do not estimate from the visible length of the answer if real records exist.

Minutes 6–9: Complete the Four-Line Token Receipt.

Record fresh input, cached input, output and multipliers. Include failed and abandoned attempts. Separate tool and search charges from model tokens.

Minutes 9–12: Calculate cost per accepted result.

Divide total workflow spend by the number of outputs that passed review. Add the human minutes needed to prepare, inspect and correct the result as a separate operating-cost line.

Minutes 12–15: Make one decision.

Choose one: cache stable context, shorten output, route a simple stage to a less expensive model, reduce retries, remove an unnecessary tool call, or accept the premium cost because it improves a consequential result.

Do not optimize seven variables at once. Change one condition, run the same job again and compare acceptance, total cost and review time. That turns a pricing opinion into evidence.

If the workflow still has no clear owner or completion standard, pause the cost optimization and use the AI readiness assessment. Cheap confusion is still confusion.

What the Lex Case Study Proves—and What It Does Not

A named use case shows the mechanism; your own usage records must prove the decision.

Anthropic publishes a named customer story about Lex, an AI-assisted writing platform. Lex reported a 50% reduction in cost after moving from a competing model to Claude, alongside changes in churn and sign-ups.

That is useful because it demonstrates the mechanism behind this article: model choice can materially change unit economics when AI is embedded in a real product. Lex did not select on token price alone. The company described evaluating tone, cost and output quality together.

The evidence has important limits. It is a provider-published customer case study, not an independent controlled trial. The public story does not provide Lex's token volumes, full cost baseline, routing rules or enough detail to reproduce the 50% figure. Product changes and other business activity could also affect the reported outcomes.

So the honest lesson is not “Claude cuts every AI bill in half.” It is this:

Lower task cost comes from matching model quality, workflow design and business acceptance—not from chasing the smallest number on a pricing page.

Use named case studies as a mechanism and a hypothesis. Use your own usage records as the decision evidence.

Before You Buy Millions of Cheap Tokens

Ben Angel helps entrepreneurs route AI by consequence and measure the result they are willing to use.

I understand why the price-per-million number is so seductive.

It turns a complicated decision into a neat comparison. One model appears ten times cheaper, so the spreadsheet seems to have made the choice for you. But the spreadsheet has not read the rejected drafts, watched the retries or sat beside you while you repair an answer that sounded right.

I have learned to ask a different question: what did I pay for the result I was willing to use?

That question forces the invisible work back into view. It makes us count failure. It stops us pretending that activity is an outcome. And it reminds us that a more capable model can be cheap when the decision matters, while a bargain model can be expensive when it multiplies supervision.

My personal rule is to route by consequence. Use the least expensive model that reliably clears the standard. Escalate when the task is ambiguous, consequential or repeatedly failing. Keep publishing, payments, deletion and customer commitments behind a human approval line.

That is the broader leadership argument inside The Wolf Is at the Door. And if you want to install the Four-Line Token Receipt inside a practical workflow, Ben Angel's 28-Day AI Mastery course helps you build one useful system with evidence, boundaries and a measurable finish line.

AI Token Cost FAQs

Rates change; cost per accepted result remains the durable buying measure.

How many words are one million tokens?

For ordinary English text, a rough estimate is about 750,000 words, but language, punctuation, formatting and code can change the ratio. Use the provider's tokenizer or actual usage metadata for a real forecast.

Does ChatGPT Plus include API tokens?

Do not assume it does. Chat product subscriptions and API billing are generally separate products. Confirm the current terms for the exact provider, plan and account before building a paid workflow.

Are output tokens more expensive than input tokens?

Often, yes. Many current model rate cards charge more for generated output than fresh input. This is why requesting unnecessary verbosity or many variants can change the bill.

What are cached tokens?

Cached input is repeated context a provider can reuse under its caching rules, often at a lower rate. Eligibility, storage duration and implementation differ by provider. Caching is useful for stable instructions and reference material, but it does not remove the need to measure total task cost.

Do reasoning tokens count?

They can. Some models perform internal reasoning that contributes to limits or billing even when all of it is not visible in the answer. Read the current model documentation and usage response rather than inferring cost from the displayed text.

How do I estimate AI token cost before launch?

Run the real task 20 to 50 times with representative inputs. Record the median and high-end input, output, tool and retry usage. Multiply by expected volume, then add a safety margin and human review cost. A single perfect demonstration is not a forecast.

Which AI model has the cheapest tokens?

The answer changes as providers release new models and discounts. More importantly, the lowest rate is not automatically the lowest cost per accepted result. Compare one real task across models and include failure, review time and tool charges.