AI Costs, Explained: Seats, Tokens, Hardware and What a Task Actually Costs

AI costs reach a small organization in three ways: a monthly fee for each person who uses a chat app (a seat), a metered charge for the text a model reads and writes through an API (tokens), or the price of hardware you buy and run yourself. Most agencies and nonprofits end up paying at least two of the three.

This guide is about product pricing for the people who use these tools, not about AI company finances. It walks through each way of paying, why the price on the pricing page is rarely what a task costs, which nonprofit and education discounts vendors publish, the AI costs that never show up on an invoice, and a simple way to estimate a monthly bill. Current prices sit in one dated table so the rest of the page stays useful after they change.

The short version: seats are predictable and cheap for light use. APIs are cheaper still for small jobs, but the bill depends on output, reasoning, caching and how many times a task calls the model. Hardware rarely pays off for a small team. And the largest of the AI costs in most budgets is staff time spent checking the output, which no vendor bills you for.

The three ways AI costs show up on an invoice

Seats. You pay a flat monthly fee per user for an app such as ChatGPT Business, Claude Team, Gemini in Google Workspace or Microsoft 365 Copilot. The fee covers a generous but limited amount of use. Heavy users hit limits, not overage charges, unless the vendor has added paid credits.

Usage-based API pricing. You pay for tokens. A token is a chunk of text, roughly three quarters of an English word. Vendors price input tokens (what you send) and output tokens (what the model writes) separately, per million. We explain the unit itself in what tokens are and the split in input vs output tokens.

Self-hosting. You buy a machine, download an open-weight model and run it. There is no per-token bill. You pay up front, then pay for power and for someone to keep it running.

A fourth pattern is growing: a seat with metered usage on top. Claude Enterprise charges a seat fee and bills usage separately at API rates. Microsoft sells Copilot Credits for its heavier agent features. OpenAI lets Business and Enterprise customers buy credits for more access. When a vendor says "seat," check whether usage is included or extra.

Per-seat subscriptions: the predictable part of AI costs

Seats are the easiest AI costs to budget. Ten people at $20 a month is $200 a month, whatever they do with it. For writing help, summaries, meeting notes and research, a seat is usually the right buy.

Three things change the math.

Annual versus monthly billing. At the vendors in our table, monthly billing costs 20% to 40% more per seat than an annual commitment. That is worth it for a pilot and not worth it once a tool sticks.

Standard versus premium seats. OpenAI and Anthropic both sell a premium seat at five times the standard price. It buys higher limits and more agent and coding time. We broke down what OpenAI’s version buys in ChatGPT Business premium vs standard. Most staff do not need one.

Bundles you already pay for. Google includes Gemini in Workspace plans, with the full set in Docs, Sheets and Meet from Business Standard up. Microsoft sells Copilot as an add-on to Microsoft 365. If your organization already pays for one of these suites, the AI may be partly paid for. The question becomes whether staff use it.

Watch for metered extras inside seat plans. Microsoft is turning usage-based billing on by default for new Copilot Business licenses, with a 4,000-credit default cap per user. We covered the settings in Copilot Credits default billing.

Usage-based API pricing: paying per million tokens

APIs are how a website chatbot, a form processor, a tagging script or a WordPress plugin talks to a model. AI costs on the API scale with use, and small jobs cost very little. A model priced at $0.10 per million input tokens reads roughly 750,000 words for a dime.

The price you see has three parts.

Input price. What you pay for everything you send: your instructions, the documents you attach, the conversation so far and the user’s question. For a chatbot that searches your knowledge base, input is usually most of the tokens.

Output price. What you pay for what the model writes. Output costs five to eight times as much as input at the major vendors. A model that writes long answers costs more than one that writes short ones, at the same list price.

Cached input price. If you send the same opening text again and again, such as a long set of instructions or a policy manual, vendors charge a fraction of the input price for the repeat. Anthropic’s cache reads cost a tenth of base input or less, depending on the model.

Every major vendor also offers batch processing at half price. You hand over a pile of requests and get results back later instead of right away. Overnight jobs such as tagging an archive, writing alt text or sorting survey answers fit this well.

Model choice matters more than vendor choice. At Anthropic and OpenAI, the largest model costs 100 times the smallest per token. Our posts on Claude Haiku 5.5 and OpenAI’s recommended migration path show how much a team saves by matching the model to the job.

Self-hosting: buying the hardware

Running a model on your own machine means no per-token bill and no data leaving the building. It also means a large check up front. A desktop box with 128GB of memory, a common size for running mid-sized open-weight models, costs thousands of dollars. NVIDIA’s store lists its DGX Spark at $6,950 as of late October 2026.

The box is only part of the AI costs of self-hosting. Add power, a place to put it, and staff time to update, patch and monitor it. Our local AI hardware prices post redid the break-even math and found that a small team needs heavy, steady use before buying beats renting. It also found hardware prices rising through 2026 while API prices held or fell.

Large open models need far more than a desktop. Our two posts on Kimi K3, renting it from a host and running it yourself, show where "open weights" stops meaning "cheap." If you want to compare machines on real work instead of spec sheets, MLPerf Inference v6.1 is the main industry benchmark for that.

For most small organizations, self-hosting makes sense for one reason: a hard rule that certain data cannot leave the building. Cost alone rarely gets you there.

Why list price is not cost per task

Two models with the same price per token can have very different AI costs for the same job. The pricing page tells you the rate. It does not tell you how many tokens a task uses.

Output length. Some models write more. Since output is the expensive part, a wordier model costs more per answer.

Reasoning tokens. Many current models think before they answer. That thinking is billed as output, even when you never see it. OpenAI’s reasoning guide says reasoning tokens are not visible through the API but are "billed as output tokens." Google labels its Gemini output price "including thinking tokens." Anthropic’s docs describe a usage field that reports how many of the billed output tokens were internal reasoning. A model that reasons at length on an easy question can cost several times what a simple answer would.

Tokenizers. Vendors split text into tokens differently, and they change it between versions. Anthropic says its Claude 4.7 and later models produce about 30% more tokens for the same text than earlier models. The per-token price can stay flat while the bill goes up.

We have measured this gap several times. Gemini 3.8 Flash cost 57% more to run than its predecessor at the same price, and Grok 4.7 kept its price while raising its cost per task. We also looked at when extra tokens help quality and when they do not, in AI token usage and quality.

Caching, retries and agent loops

Three more factors push real AI costs away from the list price, in both directions.

Caching lowers it. If most of every request is the same opening text, caching can cut the cost of the repeated part by 90% or more. How much depends on your cache hit rate: the share of input that is a repeat the vendor still has stored. Our Claude Fable 5.1 cost analysis showed the same model looking cheap or expensive depending on that one number. At Anthropic, writing to the cache costs more than normal input, so caching pays off only when the text repeats.

Retries raise it. A request that times out, returns bad JSON or fails a check gets sent again. You pay for both attempts. A workflow with a 10% retry rate costs about 10% more than the math says.

Agent loops raise it a lot. An agent does not call the model once. It plans, calls a tool, reads the result, calls the model again and repeats. Each step resends the conversation so far as input. A task that takes eight steps can use ten or more times the input tokens of a single question. Tools add to it too: Anthropic’s pricing page notes that tool definitions and a built-in system prompt add input tokens to every request that uses tools.

The practical lesson: measure cost on real tasks, not on a single test prompt. The major vendors’ API responses report token counts. Log them for a week before you budget.

Nonprofit and education discounts vendors publish

Discounts cut AI costs for seats by more than half at three of the big four vendors. They are rare on APIs.

OpenAI. OpenAI for Nonprofits prices ChatGPT Business standard seats at $8 per user per month billed annually or $10 monthly, verified through Goodstack. Premium seats are not discounted. Larger nonprofits can ask sales about up to 75% off ChatGPT Enterprise. OpenAI’s help article says the discount does not apply to the API. Verified U.S. K-12 educators can use ChatGPT for Teachers free through June 2027, according to OpenAI’s pricing page.

Anthropic. Claude’s nonprofit Team rate is $8 per user per month, and $3 for nonprofits in low- and middle-income countries. Eligibility covers 501(c)(3) organizations and international equivalents, K-12 schools and some mission-based healthcare providers. Teams under 20 people verify through Goodstack. One detail many miss: Team plans, including the nonprofit Team plan, come with monthly API credits of $20 per standard seat, pooled and capped at $500. That can cover a small chatbot or tagging script.

Google. Google Workspace for Nonprofits is free and includes basic Gemini app access. Gemini across Docs, Sheets, Meet and Drive starts at Business Standard, which nonprofits get at a steep discount. Enterprise editions are listed at 70% or more off.

Microsoft. Microsoft lists Microsoft 365 Copilot at 15% off for eligible nonprofits. We covered how that stacks with partner promotions in nonprofit Copilot discounts.

Exact rates are in the table below.

AI costs as of November 2026

Checked on each vendor’s pricing page on October 21, 2026. Seat prices are per user per month. API prices are per million tokens, input and output, with the cached-input rate.

The hidden AI costs: staff time, review and integration

The invoice is the smallest part of most AI budgets. Three AI costs never appear on it.

Review. Someone has to check what the model writes before it reaches a member, a donor or the public. If a draft takes 15 minutes to check, that check costs more than every token in it. Count review minutes per task. They decide whether a tool saves time at all.

Integration. Connecting a model to your CMS, AMS, CRM or help desk takes developer time. It also takes upkeep. Vendors retire models, change APIs and change defaults, and each change means testing. Budget a few hours each quarter for every integration you run.

Setup, training and administration. Someone sets up accounts, writes usage rules, reviews data settings, trains staff and answers questions. For seat plans, a tool nobody uses is the most common waste. Unused seats are a cost too.

There are smaller ones: security review of a new vendor, legal review of terms, and time spent comparing tools. Count them all before you compare the cost with the time saved.

A plain rule: if a task saves less staff time than it takes to set up, check and maintain, it is not saving money, however low the token price.

A simple way to estimate a monthly AI bill

You can estimate the AI costs of a single API job in five steps. Here is a worked example for a member-services chatbot.

1. Count tasks per month. The chatbot answers 3,000 questions a month.
2. Estimate tokens per task. Each question sends about 4,000 input tokens: instructions, retrieved help articles and the question. Each answer is about 600 output tokens.
3. Multiply. That is 12 million input tokens and 1.8 million output tokens a month.
4. Apply the price. At $2 input and $10 output per million, input costs $24 and output costs $18. Total: $42.
5. Adjust for real behavior. Add reasoning, retries and steps, then subtract caching.

Now adjust. If the model reasons for 1,000 tokens per answer, output rises from 1.8 million to 4.8 million tokens. Output now costs $48, and the bill is $72.

If 3,000 of the 4,000 input tokens are the same instructions each time, caching helps. At a $0.10 cached rate, those 9 million tokens cost $0.90. The other 3 million cost $6. Input drops from $24 to about $7, and the bill falls to about $55.

Finally, add a margin of 25% to 50% for retries, testing and growth. Call it $70 to $80 a month.

Two checks finish the estimate. First, run the same numbers on a smaller model. At a tenth of the price, the same chatbot costs under $10, if the answers hold up. Second, check whether you already have credits. Ten Claude Team standard seats come with $200 a month of API credit.

For seats, estimating AI costs is simpler: people times price per seat, at the annual rate if you commit. Then count how many of those people used the tool last month.

What to do before you buy

A short list for an agency, nonprofit or small publisher deciding where AI costs should go.

• Start with seats for people. Give staff one chat app, on monthly billing, for a pilot. Move to annual once usage holds.
• Use the API for repeat jobs. Chatbots, tagging, alt text and form processing are cheaper by the token than by the seat.
• Apply for nonprofit pricing first. The seat discounts above are 60% or more at three of the four big vendors.
• Pick the smallest model that passes. Test the cheap tier before the expensive one, on real tasks.
• Turn on caching and batch where they fit. They are the two largest discounts on the API side.
• Set spending caps. Every API console and most admin centers let you set a monthly limit. Use it.
• Log tokens per task for a week. Budget from that, not from the pricing page.
• Count review time. It decides whether the tool pays for itself.
• Re-check prices every quarter. They move often, and the dated table above will age.

Frequently Asked Questions