What is an LLM? How large language models work, and what they cost to run

An LLM, or large language model, is a program trained on enormous amounts of text to predict what comes next. That one trick, repeated, is what lets it answer questions, write code and summarise documents. It is the engine inside ChatGPT, Claude and Gemini, not a product you buy on its own.

If you arrived looking for the university qualification, an LLM there is a Master of Laws. The rest of this page is about the AI meaning.

What an LLM actually is

Strip the marketing and an LLM is a very large statistical model of language. During training it reads text and repeatedly guesses the next piece of a sentence, adjusting itself each time it guesses wrong. Do that across a large enough body of text, with a large enough model, and something unexpected happens: the ability to continue text turns into the ability to answer, translate, explain and write code, because all of those are special cases of “what should come next here”.

Two words you will meet immediately:

• Tokens. Models do not read words, they read tokens, which are word fragments. Pricing and limits are quoted in tokens, and a rough rule is that 1,000 tokens is about 750 English words.
• Context window. How much text the model can hold in view at once, counted in tokens. A large window is what lets you paste a long contract and ask questions about it.

How an LLM works, in the order it happens

1. Your text is split into tokens.
2. Each token becomes numbers, a representation capturing how it is used relative to other words.
3. The model weighs every token against every other one to work out what matters in this particular sentence. This is the attention mechanism, and it is the reason modern models handle long passages far better than earlier ones.
4. It produces a probability for every possible next token and picks one.
5. It repeats, now including what it just wrote, until the answer is finished.

That last step explains the single most important property of these systems. The model is predicting plausible text, not retrieving verified facts. When it does not know something, nothing in the mechanism makes it stop; it produces the most plausible-looking continuation instead. That is where confident wrong answers come from, and no amount of paying for a higher tier removes it.

LLM examples, and what you can actually buy

The models most people name are the ones inside consumer assistants. In our catalogue, the model layer is its own large language models category, with 11 active listings including OpenAI, Google Gemini API, xAI Grok, DeepSeek, Cohere, Alibaba Qwen and IBM Granite. None of those listings carries user reviews yet, so there is no rating to quote for any of them.

That is worth noticing rather than glossing over. Reviews accumulate for products people administer and renew. The model layer is bought by developers through an API, so it generates far less review activity than the assistant sitting on top of it, which is why the consumer products like Claude have ratings and the models underneath mostly do not.

LLMs versus chatbots: what the difference actually is

An LLM is the engine. A chatbot is the car.

The model takes text and returns text. A chatbot wraps it in everything that makes it usable: a conversation interface, memory of what you said earlier, guardrails, the ability to search the web or read your files, and an account to bill. Swap the model inside a chatbot and it is the same product with different performance. Strip the chatbot away and the model cannot even remember your last message, because the memory lives in the wrapper.

This is why “which LLM is best” is usually the wrong question for a buyer. You are choosing a product, and the product’s context handling, integrations and terms will matter more to your week than which model is two points ahead on a benchmark this month.

What an LLM is in ChatGPT

ChatGPT is the product; the LLM is the model it calls. One subscription typically gives access to several models of different sizes, and the product routes between them depending on the question and on how busy the service is. That routing is why the free tier can feel inconsistent: at quiet times you may get a stronger model than at peak.

The same split holds across the market. You subscribe to Claude, Gemini or ChatGPT; what you are actually renting is metered access to models underneath.

What LLMs cost

Two completely different pricing worlds, and comparing a figure from one against the other tells you nothing.

If you subscribe, you pay per person per month. Claude Pro and ChatGPT Plus are both $20 billed monthly, and Google AI Pro is $19.99. Our Claude Pro price and ChatGPT Plus price breakdowns cover those ladders properly.

If you build, you pay per million tokens, split between what you send and what comes back. From Anthropic’s own published rates in US dollars:

Model	Input per 1M tokens	Output per 1M tokens
Fable 5.1	$10	$50
Opus 5.5	$4	$20
Sonnet 5.5	$2	$10

For contrast at the cheap end, third-party trackers put DeepSeek’s Flash model near $0.15 input and $0.60 output per million tokens off peak. That is a spread of more than thirty times on output across the market, which is why picking a model for the job rather than always reaching for the strongest one is the main lever on an AI bill.

Three cost facts worth carrying:

• Output costs several times more than input on every vendor. Asking for brevity genuinely reduces the bill.
• Caching and batching cut it materially. Anthropic publishes batch processing at half price.
• A bigger model is not a better answer for simple work. Classification and extraction rarely need the expensive option.

Frequently asked questions

Related reading

• What is an AI agent?: what gets built on top of these models.
• Claude Pro price: subscription pricing, verified from Anthropic’s own page.
• ChatGPT Plus price: the same ladder on the OpenAI side.
• Claude vs ChatGPT: choosing between the two products built on these models.
• Large language models on Spotsaas: all 11 listings in the category.