Quick answer
AI inference is a trained model producing an answer from your input, such as 1 question. On the two developer price pages read here (checked October 7, 2026), the billing units are minutes and searches: OpenAI voice at $0.05 per minute and Anthropic web search at $10 per 1,000 searches. Per-use prices count units of use.
Key takeaways
- Inference is a trained model producing an answer from new input. Google Cloud, Cloudflare and IBM each describe it as the model being put to work after training.
- Cloudflare calls training more or less a one-time expense and inference ongoing. Its point is that if a model is actively in use, it keeps applying its training to new data.
- The two developer price pages read for this piece (OpenAI and Anthropic, checked October 7, 2026) state different units: OpenAI voice sessions at $0.05 per minute and Anthropic web search at $10 per 1,000 searches; the table shows how each page labels its currency. They are different products and units, not a price comparison.
- If your tool is a flat monthly plan, this page's prices are not your bill. If it is billed per use, estimate your monthly volume before you sign.
- Read the plan page of the app you use before relying on any per-use price; the worked lines here only show the method.
What is inference in AI?
AI inference is a trained model taking your input, such as 1 question or 1 phone call, and producing an answer. Cloudflare's idea is that if a model is actively in use, it "is constantly applying its training to new data". This site's reading: an owner meets the word because prices count use. Training is the earlier step that builds the model. Think of a cook: culinary school is training, and cooking each order that comes in is inference.
Three vendor explainers describe it in similar terms. Google Cloud calls AI inference "the 'doing' part of artificial intelligence". Cloudflare says "inference is an AI model in action". IBM defines it as "the act of using a trained AI model to make predictions on new data".
| Source (read October 7, 2026) | How it describes inference | What it says about cost |
|---|---|---|
| Google Cloud | "the 'doing' part of artificial intelligence" | Each prediction is "far less computationally demanding than training", but delivering millions of predictions in real time requires highly optimized, scalable infrastructure |
| Cloudflare | "an AI model in action" | Training "can be very expensive in terms of compute power" but is "more or less a one-time expense"; "Inference, however, is ongoing" and "can be very expensive" |
| IBM | "the act of using a trained AI model to make predictions on new data" | The goal of inference is deploying the trained model in a "maximally efficient, cost-effective way" |
The cost column is not a single story: Google is comparing one prediction with training, Cloudflare is describing when each expense recurs, and IBM is describing a goal. None of the three pages gives a price for you, and none says which costs more over a model's whole life. This page translates them.
What am I paying for when I use an AI tool?
On a per-use plan you pay for units of use, such as a minute or a search; the two examples below are a session fee and a tool fee, and both vendors say other usage is charged separately or on top. OpenAI's page says "Backend model and tool usage is charged separately", and Anthropic's says web search comes "plus standard token costs for search-generated content".
This site's reading, not any page's: Cloudflare says a model in use "is constantly applying its training to new data and making additional inferences", so inference happens every time the model is used, and a price that counts use is counting something that recurs in the same way. None of the explainers says a billing unit is inference. IBM adds that loading a model's parameters into memory each time inference is performed "entails energy usage and costs".
Why does inference show up on AI price lists?
On the two price pages read here, the billing units are minutes and searches; the word "inference" appears only in OpenAI's sentence "Inference discounts are available" (if you enable data sharing when creating a fine-tune job) and in Anthropic's "US-only inference" setting. Per-use prices count units of use. Both pages are aimed at developers who call a model from their own software, not at owners buying an app.
| Vendor price page (developer prices, checked October 7, 2026) | Unit and amount as the page shows it | Currency | Not in that price | This site's worked line |
|---|---|---|---|---|
| OpenAI API pricing, "GPT-Live sessions", gpt-live-1 voice | "$0.05" under "Price per minute"; "billed per second, without rounding up to a whole minute" | Shows a dollar sign; the page does not name the currency | "Backend model and tool usage is charged separately" | 300 minutes x $0.05 = $15 in the page's unnamed currency |
| Anthropic pricing, web search on the Claude API | "$10 per 1,000 searches" | "All prices are in USD" | "Plus standard token costs for search-generated content" | 300 searches x $10 / 1,000 = $3 |
These are different products and units, a voice session and a search tool, so the two lines are not a price comparison between the vendors.
Read the worked lines with these limits:
- They are this site's arithmetic on volumes we picked (300 minutes, 300 searches) to show the method, not typical costs.
- They are developer prices, not an app bill.
- They are incomplete, because other usage is charged on top (backend and tool usage for OpenAI, token costs for Anthropic).
Both vendors also bill for tokens (the chunks of text a model reads and writes), which is a separate subject.
The word itself is rare on these pages. OpenAI's pricing page says "Inference discounts are available" if you enable data sharing when creating a fine-tune job, and Anthropic's mentions a "US-only inference" setting that changes token pricing for some Claude models. That both use the word for a model handling a request is this site's reading, not either page's.
Do I need to know this?
Only if your tool bills per use; for a flat monthly plan, this site's rule of thumb is to ask what the plan covers instead. OpenAI's help center says "API usage is billed separately from your ChatGPT subscription", and Claude's says "A paid Claude subscription enhances your chat experience, and the Claude API is billed separately." Each describes its own developer billing next to its own subscription. How a third-party app prices itself is not stated on the pages we read.
With a per-use plan, the unit tells you what the meter is counting. Two numbers matter: the unit price on the vendor's page, and how many units you will use in a month. The second number is yours, not the vendor's.
How do I estimate a per-use AI bill?
Multiply your monthly volume by the unit price, then check what the price leaves out. This procedure is this site's advice, not a vendor's.
- Find the billing unit on the price page, such as per minute, per search or per token.
- Count your own monthly volume in that unit from a phone log, an inbox or last month's work.
- Multiply volume by the unit price. Do it for a quiet month and a busy one.
- Read what the price excludes. In the two examples above, the vendors say tool usage or token costs are charged separately.
- Write each amount down as the page shows it, such as $0.05 per minute or $10 per 1,000 searches, and check the table above for how each page labels it.
For per-call bills worked at several volumes, see AI Receptionist Cost. For another per-use meter, see why Zapier bills the way it does. The wider picture is in the practical guide to AI automation for small business.
What should I do next?
Before you sign for any AI tool, ask the seller these five questions: three come from what the price pages above show, and two are this site's own.
- Is the price per second, per minute, per call or flat? (OpenAI's voice page bills per second against a per-minute price; Anthropic's search is priced per 1,000 searches.)
- Are tools, search or backend usage charged on top? (Both pages say so.)
- In which currency is the price? (One page says USD; the other shows only a dollar sign.)
- What happens past the included volume? (This site's own; neither page covers it.)
- Is there a cap or an alert before the bill grows? (This site's own; neither page covers it.)
If the plan is flat, check its page for usage limits. If it is per use, run the five steps above with your own volume. Keep the vendor's page open while you do it; prices change, and the ones on this page were checked on October 7, 2026, from where this page was read.
Disclosure: this page was drafted with Claude, a model made by Anthropic, one of the vendors whose price page is read here.
FAQ
Is ChatGPT inference?
ChatGPT is OpenAI's chat product. When a trained language model answers your message, that is inference as the explainers define it: Cloudflare's lists large language models replying to new prompts among its common use cases, and says almost any real-world application of AI relies on inference.
What is the difference between inference and generative AI?
Generative AI is one kind of machine learning; inference is the step of using a trained model. IBM's explainer says all machine learning, including generative AI, boils down to pattern recognition, and that any instance of an AI model generating outputs or making decisions in a real-world application constitutes inference.
What is the difference between inference and training?
Training builds the model and inference uses it. Cloudflare describes training as the first phase and inference as the process that follows. Google Cloud says each individual prediction is far less computationally demanding than training, but that delivering millions of predictions in real time requires highly optimized, scalable infrastructure.
Do I pay for inference if I have a flat monthly AI plan?
The two vendor help pages we read say their API usage is billed separately from their own subscriptions: OpenAI's says API usage is billed separately from a ChatGPT subscription, and Anthropic's says the Claude API is billed separately from a paid Claude subscription. For any other app, read its own plan page.
Why do some AI price lists say inference?
Only some do. OpenAI's developer pricing page says inference discounts are available if you enable data sharing when creating a fine-tune job, and Anthropic's mentions a US-only inference setting that changes token pricing for some Claude models. This site's reading: both use the word for a model handling a request.