AI models explained: names, prices and how to choose

Prices and picks updated from our catalog on 17 October 2026 · Text revised 18 October 2026

Ask GPT-6 Luna and GPT-6 Astra to summarise the same document, and one bill is $0.00045 while the other is $0.045: 100 times more. Which one you should use depends on the job, and names like Claude Haiku 5.5, GPT-6 Luna, and Gemini 4 Argon don't tell you much until you know the pattern. This guide decodes the names, shows how the bill is worked out, and ends with picks you can test side by side in one click.

What is an AI model?

An AI model, often called an LLM (large language model), is the program behind a chatbot. You send it text, and sometimes images or files, and it writes an answer. ChatGPT, Claude and Gemini are apps; the models are the engines inside them, and the same engines are sold to developers, who pay for each request.

The companies that build them are called labs. All models do the same basic job, but they differ in four ways that matter to you: how good the answers are, how much they cost, how fast they reply, and what they can read (text only, or also images, PDFs, audio and video).

Why AI models come in sizes (Haiku vs Sonnet vs Opus)

Most labs sell a family of models in sizes, like small, medium and large coffee. The small one is fast and cheap and handles everyday jobs well. The large one is slower and costs more, and it is usually better at hard problems such as tricky code, maths or long chains of reasoning. The name tells you the size: in the table below, Haiku, Luna, or Flash Lite mean small, and Opus, Astra, or Argon mean large.

Prices are US dollars per 1 million tokens of input and output. A higher price doesn't guarantee a higher score: Claude Fable 5.1 scores 53.4 on the intelligence index, below Claude Opus 5.5 at 57.6, for 2.5 times the price.

The number in the name is the version. Compare numbers only within one size: a higher number means a newer model of that size, not a bigger one. Labs retire old versions over time, so it rarely pays to use an old one. Some labs sell one main model rather than a range, such as xAI with Grok 4.7 and Meta with Muse Spark Open.

You will also see open-weights models. Anyone can download and run them, and many companies host them, which often keeps prices low. 18 of the 30 models in ModelCompare's catalog are open, mostly from Alibaba, DeepSeek, Kimi, Mistral AI, and Xiaomi.

Test Claude Haiku 5.5, Claude Sonnet 5.5, and Claude Opus 5.5 on your own prompt

How AI model pricing works: tokens, input and output

AI models charge by the token, a chunk of text about three quarters of an English word long. A thousand tokens is roughly 750 words, and prices are quoted per million tokens, which is about 750,000 words or ten novels. Text in languages such as Chinese, Arabic, Hindi or Russian often takes more tokens per word than English, so the same message can cost more.

Every request has two parts with two prices. Input is everything you send: your question, any document you paste and the earlier messages in the chat. Output is what the model writes back. Output costs more, about 4 times the input price for a typical model in our catalog. Claude Haiku 5.5, for example, charges $0.10 per million input tokens and $0.50 per million output tokens.

A worked example

Say you paste a 1,500-word document (about 2,000 tokens) and ask for a 375-word summary (about 500 tokens). Here is what one summary costs at list prices, and what 1,000 of them would cost:

The same job costs about 100 times more on GPT-6 Astra than on GPT-6 Luna. For one summary that is small change; for an app sending thousands of requests a day, it decides whether the product makes money. These figures assume the model doesn't think before answering. Real bills are often higher, as the next section shows. Try your own document on all 6

Four things that change the bill

• Hidden thinking. Many models, even small ones, think before they answer, and those thinking tokens are billed as output even though you never see them. In our 241-prompt cost test of four cheap models (October 2026), Claude Haiku 5.5 and GPT-6 Luna had the same list price, yet Haiku's bill was 1.7 times Luna's because half its output was thinking. Gemini 3.8 Flash cost 7.5 times Luna per token but 18 times more on the same work. The price per token is not the price per task.
• Long chats. Each new message resends the whole conversation as input, so a long chat gets more expensive with every turn. Start a fresh chat when you change topic.
• Cached input. Many models charge much less for input they have seen recently, such as the same long document asked about several times. This matters mostly for developers.
• Language. Non-English text usually needs more tokens for the same meaning, so it costs more.

Monthly subscription or pay per token?

Chat apps sell a flat monthly plan with usage limits. Developers, and tools like ModelCompare, pay per token instead. Pay-per-token is usually cheaper for light or occasional use: 1,000 of the summaries above cost $0.45 on GPT-6 Luna. A plan tends to win if you chat with a top model for hours every day.

Context window, speed and other specs explained

Intelligence score

The intelligence score is a single number from the Artificial Analysis Intelligence Index, which combines reasoning, knowledge, maths and coding tests. Higher is smarter. Scores are averages of public tests: models can be tuned to them, and they say little about tone, formatting or your niche, so treat a gap of a few points as a tie.

Context window

The context window is how much text a model can read at once, counting the whole chat. 22 of the 30 models in ModelCompare's catalog take a million tokens or more, roughly 750,000 words or several books.

Speed

Speed is how long you wait for an answer. In our October 2026 test the median answer took 1.4 seconds on Claude Haiku 5.5, 1.6 seconds on GPT-6 Luna and 5.8 seconds on Gemini 3.8 Flash, which spent longer thinking.

Knowledge cutoff

The knowledge cutoff is the date a model's training data ends. A model doesn't know about later events unless you paste them in.

Inputs

Inputs are what a model can read besides text. Most read images and PDFs; a few also take audio and video. These are called multimodal models.

How to choose the right AI model

1. Start small. Try your task on a small, cheap model first. For emails, summaries, translation and everyday questions it is often just as good, and faster.
2. Move up only when it fails. If the answer is wrong or shallow, try the next size up in the same family, then a top model or a reasoning setting.
3. Test on your own prompt. Rankings average many tests, and your task is not the average. Send the same prompt to a few models side by side and compare the answers and what each one cost.
4. Weigh cost against volume. For a one-off question, pick the best answer. For something you'll run thousands of times, a model that is 90% as good for a tenth of the price usually wins.

Here are some starting points, each one click away from a side-by-side test. Prices are blended per million tokens; scores are from the intelligence index.

Still between two? Head-to-head pages are often the quickest way to decide: Compare Claude Haiku 5.5 and GPT-6 Luna , Compare Claude Sonnet 5.5 and GPT-6.1 Sol or Compare Claude Opus 5.5 and Gemini 4 Argon .

Frequently asked questions

Glossary

Token

A token is a chunk of text about three quarters of an English word long. AI models read, write and bill in tokens.

Prompt

A prompt is what you send to an AI model: your question, instructions and any text you paste.

Reasoning model

A reasoning model thinks step by step before it answers. The thinking improves maths and code but is billed as output and adds waiting time. Many models let you turn it up or down with an effort setting.

Context window

A context window is the most text an AI model can read at once, counting the whole conversation so far.

Open weights

An open-weights model can be downloaded and run by anyone, so many companies host it. Check its licence: open weights does not always mean free for any use.

Multimodal

A multimodal model reads more than text, such as images, PDFs, audio or video.

Hallucination

A hallucination is a confident answer that is wrong or made up. Every AI model does it sometimes, so check facts that matter.