What people use this for
- Budgeting an AI project before you write a proposal
- Deciding between per-seat subscriptions and metered API access
- Working out which model tier a workload actually needs
- Sanity-checking a quote from a vendor or an agency
Almost every AI cost question comes down to the same confusion: providers bill in tokens, businesses budget in months, and nobody publishes the conversion. The calculator above does that conversion. This page explains what it is doing, so the number means something when you take it to whoever signs the cheque.
How AI billing actually works
There are two ways to pay for AI, and they behave nothing alike. A per-seat subscription is a flat monthly fee per person for a chat interface — predictable, capped, and largely indifferent to how hard someone uses it. API access is metered: you pay for each unit of text going in and each unit coming out, which means your bill tracks usage exactly and starts at effectively zero.
The unit of metering is the token — roughly three-quarters of a word in ordinary English. Every provider charges separately for input tokens (what you send: the prompt, the pasted document, the conversation so far) and output tokens (what the model writes back). Output is always the more expensive of the two, at every provider, usually by a multiple rather than a margin.
Why conversations cost more than they look like they should
Models have no memory between calls. To continue a conversation, the entire history is re-sent on every turn — so turn ten pays for turns one through nine again. A twenty-message exchange does not cost twenty messages, it costs something closer to the sum of a growing series, and the last few turns dominate.
This is the single most common reason a real bill exceeds an estimate. It is also why agent workloads — the ones that take several steps, call tools, and accumulate context as they go — are dramatically more expensive per task than a single question, and why the calculator above treats them as a separate workload rather than a bigger version of a chat message.
Subscriptions or API: which is cheaper for you
The honest version of a comparison that usually gets sold rather than explained.
Per-seat subscriptions win when
- People use it heavily, every day, at a steady rate
- You want a fixed number in the budget with no variance
- Your team needs the polished interface, not a raw API
- Nobody on staff is going to build anything
- You would rather not think about usage at all
Metered API access wins when
- Usage is light, spiky, or concentrated in a few people
- The work is automated rather than typed by a person
- You want to route easy tasks to a cheap model
- You are embedding AI into your own software
- You need cost per process, not cost per person
Most businesses that have both end up with a split: subscriptions for the handful of people who live in the tool all day, API access for everything that runs without a person watching. That is usually cheaper than committing to either, and it is what the crossover figure in the calculator is pointing at.
Paying for capability you do not need
The gap between the cheapest and most expensive current models is roughly an order of magnitude — but the gap in capability on a well-defined task is often close to nothing. Classifying an email, extracting a date from an invoice, or routing a support ticket does not need a frontier model, and running one for those jobs is the most common avoidable line on an AI bill.
A sequence that reliably lands on the right model without guesswork:
Start on the balanced tier
It handles the overwhelming majority of business work. Starting at the frontier tier teaches you nothing about what the task needs and costs several times more while you find out.
Try the cheap tier on the same task
Run both against twenty real examples and compare. On narrow, well-specified jobs the cheap model frequently matches the balanced one — and when it does, that is most of your bill gone.
Escalate only where quality actually fails
Move to the frontier tier for the specific tasks that fail on cheaper models, not for the whole workload. Mixed routing is normal and is how experienced teams run.
Re-test when models change
Capability moves down the price curve constantly. A task that needed the frontier tier last year often runs fine on the balanced one now, and nobody sends you an email about it.
What the calculator deliberately leaves out
Token cost is usually not the largest number in an AI project, and a calculator that implied otherwise would be misleading. Set against the monthly figure above, budget for the work of integrating it, the ongoing cost of someone checking output quality, and the retries — failed generations, rejected drafts, and second attempts all bill normally.
The costs that do not appear in any token count:
- Building and integrating the thing, which usually dwarfs the running cost in year one
- Human review time, which is a permanent line item and not an optional one
- Retries and rejected output — every attempt bills, including the bad ones
- Prompt iteration during development, which burns real tokens on throwaway work
- The management overhead of a system somebody has to own
How much does AI cost for a small business per month?
For a team of five doing ordinary drafting and summarising work through the API, monthly costs typically land in the low tens of dollars — genuinely less than most software subscriptions. Per-seat subscriptions for the same team cost more but are fixed and predictable. Costs rise sharply with agent workloads and long documents, because both re-send large amounts of context on every step. Use the calculator above with your own numbers rather than a benchmark from someone else.
Is the AI API cheaper than a ChatGPT or Claude subscription?
It depends entirely on usage intensity. For light or occasional users the API is dramatically cheaper, often by an order of magnitude, because you pay only for what you use. For heavy daily users the subscription usually wins, since it caps out while metered billing keeps climbing. The calculator names the specific crossover point for your team size and workload.
What is a token and why am I billed for them?
A token is a chunk of text — roughly three-quarters of a word in English, though common words are one token and rare ones split into several. Models process text as tokens, so providers bill by them. Input tokens (what you send) and output tokens (what the model writes) are priced separately, with output always costing more.
Why is my actual AI bill higher than I estimated?
Almost always one of three things. Conversation history is re-sent on every turn, so long exchanges cost far more than the number of messages suggests. System prompts and pasted context are billed on every single call, not once. And failed or rejected generations bill the same as successful ones. Agent workloads combine all three, which is why they are the most common source of bill shock.
How can I reduce my AI costs?
In order of impact: route simple tasks to cheaper models rather than running everything on a frontier model; cap conversation history instead of letting it grow unbounded; ask for shorter outputs, since output tokens are the expensive ones; use prompt caching where your provider offers it, which cuts the cost of repeated context substantially; and batch anything that is not time-sensitive, which most providers discount heavily.
The short version
- Output tokens cost several times more than input tokens — the shape of a workload matters more than its size.
- Conversations re-send their whole history every turn, which is the usual reason a bill beats its estimate.
- Start on the balanced tier, test the cheap tier on the same task, and escalate only where quality actually fails.
- Subscriptions win for heavy steady users; metered API wins for light, spiky, or automated work. Most businesses end up with both.
- Token cost is rarely the biggest number — integration and human review usually are.
Current model prices
USD per million tokens, standard tier. Batch, cached-read, and committed-spend discounts are excluded — they vary by account and would make the comparison flattering rather than useful.
| Model | Tier | Input | Output | Context |
|---|---|---|---|---|
| Claude Opus 4.1Anthropic | Frontier | $15 | $75 | 200K |
| Claude Sonnet 4.5Anthropic | Balanced | $3 | $15 | 200K |
| Claude Haiku 4.5Anthropic | Fast & cheap | $1 | $5 | 200K |
| GPT-5OpenAI | Frontier | $1.25 | $10 | 400K |
| GPT-5 miniOpenAI | Fast & cheap | $0.25 | $2 | 400K |
| Gemini 2.5 ProGoogle | Frontier | $1.25 | $10 | 1M |
| Gemini 2.5 FlashGoogle | Fast & cheap | $0.3 | $2.5 | 1M |
Output tokens cost several times more than input tokens at every provider, which is why a job that reads a lot and writes a little is far cheaper than the headline rate suggests — and why "make it more concise" is a cost control, not just a style note.