How much does AI cost? A chat, an API call and an agent

AI costs most people nothing or about $20 a month: that is the price of the paid chat plans from OpenAI, Anthropic and Google, and the free plans cost $0. Through an API you pay per token, and one ordinary question costs from about $0.0004 on a small model to $0.035 on the largest. An AI agent makes many calls for one task, so a task costs from a few cents to a few dollars, set mostly by how long its conversation grows.

Those are 3 different bills, and the right one depends on who uses the AI and how. The prices below are the makers' list prices in US dollars, read in October 2026. For a figure on today's prices, the cost calculator works out any model, and /prices lists every one.

What does ChatGPT, Claude or Gemini cost a month?

A paid chat plan costs about $20 a month at each of the 3 large makers, with a free plan below it and plans of $100 and more above it for people who use AI all day.

A plan's price is fixed; what it limits is use. ChatGPT Plus "may include usage limits such as message caps", Claude's plans list usage limits, and Google sells AI Plus and AI Pro as 2 and 4 times the usage of Free, and AI Ultra as 5 or 20 times AI Pro's. When you reach a limit you wait for it to reset or, on some plans, buy credits for more. The plans do not include the API: OpenAI says plainly that API use "is separate and billed independently".

How much does one question cost through the API?

Through an API you pay for the text the model reads (input tokens) and writes (output tokens), each priced per million TokenTokenThe small piece of text a language model reads and writes, often part of a word. Prices, limits and the context window are all counted in tokens.s. A token is about three quarters of an English word, so 100 tokens are about 75 words, by OpenAI's rule of thumb.

Take a question with some context, 1,000 tokens in, and an answer of 500 tokens out, roughly 750 words and 375:

The sum for Sonnet 5.5 is 1,000 × $2 / 1,000,000 + 500 × $10 / 1,000,000 = $0.002 + $0.005 = $0.007. Two things follow from the table. The largest model costs 100 times the smallest for the same question, so the model you pick moves the bill more than anything else. And output costs 5 times input on every model here, so a long answer costs more than a long question.

Gemini 4 Argon's price doubles on 1 December 2026 to $4 and $20, the same as Opus 5.5; the comparison of the 3 flagships works through what that does to a month.

Is a $20 plan cheaper than paying per token?

For someone who chats a lot, the plan usually is; for light use, the API costs less. The reason is that a chat is not one question. Each new message sends the whole conversation so far back to the model, so a conversation's later turns cost more than its first.

Take a 10-turn conversation that starts with 1,000 tokens of instructions and adds a 150-token question and a 400-token answer each turn. The input of turn 1 is 1,150 tokens, turn 2 is 1,700, and turn 10 is 6,100: 36,250 input tokens and 4,000 output tokens in all.

So 3 longish conversations a day on Opus 5.5 cost about what a $20 plan costs, before any caching. Ten a day would cost over $60 through the API, while the plan stays at $20 until its limit. The plan also brings the apps and their tools, which you would build yourself on the API: Claude's plans include web search and memory, and ChatGPT Plus includes image generation and file uploads.

The API wins when the work is not a person chatting: summarising a thousand support tickets, tagging products, pulling fields out of invoices. That work runs on a small model at a fraction of a cent a call, and batch processing halves it again: Anthropic and OpenAI both charge half for requests sent as a batch.

What does an AI agent cost per task?

An AgentAgentA model that uses tools in a loop, deciding each next step from what the last one returned, until the task is done. costs the sum of its steps, and every step resends the conversation, so a task's cost grows faster than its step count. An agent that reads files, calls tools and writes code adds each result to its context, and the next step reads all of it again.

Take a 20-step task that starts with 10,000 tokens (instructions, tool definitions, the request) and adds 2,000 tokens a step, with 500 tokens of output each step. Its context ends at 48,000 tokens, but the 20 steps send 580,000 input tokens in all.

On Claude Sonnet 5.5 that is 580,000 × $2 / 1,000,000 = $1.16 of input plus $0.10 of output: $1.26 a task. With Prompt CachingPrompt CachingLetting the API keep the processed start of a prompt for a few minutes, so later requests that begin the same way are cheaper and quicker., each step reads the part it sent before from the cache at $0.10 per million instead of $2, and writes only the new 2,000 tokens at $2.50 per million. The same task then costs about $0.27: $0.05 of cache reads, $0.12 of cache writes and $0.10 of output.

Caching cuts this agent's bill by nearly four fifths, which is why it is the first thing to turn on. The next is to keep tasks short, since step 20 costs far more than step 1. What an AI agent costs per month prices a support bot, a research agent and a coding session the same way, and the agent cost simulator runs your own steps on any model.

Which AI platform has predictable costs?

A subscription or seat is the predictable bill: the price is fixed and the plan limits your use instead. API use is metered, so its bill moves with traffic, but you can cap it.

On the Claude API, each of the Start, Build and Scale tiers has a monthly spend cap, and you can set your own limit below it for the whole organisation or for one workspace. When use reaches the limit, requests are refused until the next month or until you raise the limit. That turns an open meter into a known worst case. How to set budget limits on an AI agent covers the same idea inside your own code, per run and per user.

Watch the plans that mix the two. Claude Enterprise is $20 a seat a month, billed yearly for at least 20 seats, plus usage at API rates, so its bill moves with use like the API's. ChatGPT Business and Pro let you buy credits once the included use runs out, which is predictable only if you leave credits off.

What else is on an AI bill besides tokens?

Tools and running time come on top of the tokens. On the Claude API, web search costs $10 per 1,000 searches, code execution $0.05 an hour per container after 50 free hours a day per organisation, and Claude Managed Agents $0.08 for each hour a session is running. A research agent that runs 5 searches a task adds 5 cents a task in searches alone, a fifth of the cached Sonnet task above.

Then the costs that are not on the model maker's bill at all: the servers and databases your app runs on, the logs you keep, and the people who build, test and check the work. They are the same whichever model you choose.

How to estimate your own AI cost

Count the tokens of one real request, multiply by the price, then multiply by how many you run.

1. Paste a real prompt and a real answer into the token counter to get the input and output tokens.
2. Price them on 2 or 3 models in the cost calculator. For a chat or an agent, count the resent conversation, or use the agent cost simulator.
3. Multiply by a month's requests, then take off what caching and batch processing save.
4. Try the smallest model that does the job well; which model to use starts from the job and shows the cheapest models that score high enough.

Cutting the bill once it is running is its own subject: how to reduce LLM API costs goes through caching, batching, smaller models and shorter prompts with their savings.