What Is AI Inference vs Training? (2026): The Bill You Actually Pay

Training is the work that produces a model. Inference is every time that model answers you. A chat subscription pays for inference. An API bill pays for inference by the token. Training is a lab cost, or a separate fine-tuning project, and it is the wrong word for “I asked ChatGPT a question.”

Prices below are US list prices excluding tax from directory pages checked October 16, 2026, published here on October 24, 2026. The unit those APIs count is what AI tokens are. The model class is what an LLM is.

The two jobs

	Training	Inference
What happens	Weights are updated from examples	A finished model produces the next tokens
Who pays, for most readers	The lab that shipped the model	You, through a seat or an API
When it runs	Before the product exists, or during a fine-tune	Every prompt, every image, every meeting summary
What a failure looks like	A model that never learned the pattern	A wrong answer, a slow answer, or a bill that grew with volume
The lever you control	Whether to fine-tune at all	The prompt, the model, and how often you call it

flowchart TD
  train["Training updates the weights"] --> model["A finished model"]
  model --> infer["Inference: one request, one answer"]
  infer --> seat["A monthly seat, if you chat"]
  infer --> api["A token bill, if you call an API"]

A prompt does not train the model. Writing a better instruction is what prompt engineering is. Changing the weights on your own examples is what fine-tuning is. Most teams never need that second project.

What you are buying in a product

Way you use a model	What the invoice is	A published price
Chat in an app	Inference, capped by the plan	Claude Pro is $20/month, or $17 billed annually. ChatGPT Plus is $20. Go is $8
A developer API	Inference, metered in tokens	DeepSeek chat is free. API input is from $0.15 per 1 million tokens off-peak on a cache miss. Output is $0.60 off-peak
A coding agent in a repo	Inference against your code, plus the editor seat	OpenAI Codex sits on the coding side of the same split. The editor comparison is Cursor vs Claude Code vs Codex
“Train it on our wiki”	Usually retrieval, sometimes fine-tuning	Retrieval is what RAG is. It does not update the weights

DeepSeek’s peak hours, recorded on the directory page, are 01:00–04:00 and 06:00–10:00 UTC on weekdays. Peak input on a cache miss is $0.30 per 1 million tokens, and peak output is $1.20. That gap is an inference schedule. It is a training decision only for the lab.

ChatGPT Pro, from $100 a month, raises inference limits inside the product. It does not hand you the weights. Claude Max, from $100 a month, is the same kind of purchase: more inference on the model Anthropic already trained.

How to talk about it at work

Say “inference” when the question is cost per answer. A support bot that replies 15,000 times a month is an inference forecast. Chatbase prices that shape as message credits, from Hobby at $32 a month billed yearly for 700 credits. The credit is the product’s name for inference plus the wrapper around it.

Say “training” when someone proposes to change the model. That proposal needs examples, an evaluation set, and a reason a prompt plus your documents failed. The site’s model table, updated October 19, 2026, lists per-million-token inference prices for hosted models. Those numbers are answers, not the cost of the original training run.

Say “fine-tuning” only for the middle case. You start from a trained model and continue training on your set. You still pay inference afterward, every time you call the result.

FAQ

Does chatting train the model on my company data?

A vendor policy can say customer chats are excluded from later training. That policy is a contract term. The chat you just sent is still inference. Read the plan you are on before you paste a client file. This article does not score those policies.

Why is the API cheap and the chat seat $20?

The seat bundles the app, the limits, and the support around inference. The API bill is closer to raw inference, and it grows when you send more tokens. A light chat user can spend less on Plus than a script that calls the API all day.

Is generating an image inference?

Yes. The model that draws the picture is already trained. Each image is an inference call, whether the product hides it inside ChatGPT Plus or bills tokens on the image API. ChatGPT’s image API rate recorded for GPT-Image-2.5 Sunburst is $8 per 1 million image input tokens and $30 per 1 million image output tokens.

When should a team train its own model?

When a measured task stays wrong after a clearer prompt and a retrieved source, and you have examples you are allowed to train on. That is a project with an owner. A new chatbot seat is the ordinary next step.

Last verified

October 16, 2026 for plan and API prices. Published October 24, 2026. The training-versus-inference split does not depend on a price change.

Sources: Claude pricing, DeepSeek API pricing, OpenAI ChatGPT pricing.

Published by Andrew for AI Tools Optimizer.

Claude and ChatGPT are in the chatbots directory. Directory pages may use affiliate links. A purchase through them may earn us a commission at no extra cost to you.