Claude, ChatGPT, Gemini, and other US and Chinese AI services compared: pricing, privacy, and how to choose

Based on: each company's official pricing, model, and privacy pages, checked in October 2026. Dollar prices are from the US pages. Won (KRW) prices are from the Korean pages.

There are so many AI services that it is hard to know where to start. But only a few things matter when you choose: how much you can do for free, whether the service is good at the job you have in mind, and where the text you type is stored and whether it is used for training.

Names, prices, and features change every few months, so check the official page once more before you pay. This guide does not rank anything. It only copies what each company says on its official pages, and where an official page gives no value, the guide says so. It explains how 12 US and Chinese AI services differ, centered on Claude, ChatGPT, and Gemini, and how to choose between them.

1. Four words to know first

Figure 1. Three ways to use one model. The price and where your text goes depend on which way you use it (diagram, labels in Korean)

Even at the same company, app pricing and API pricing are separate. The app is a subscription with a fixed monthly fee. The API is counted in tokens. A token is the unit AI uses to count text, a piece slightly smaller than a word. "Dollars per 1 million tokens" on an API price list is the cost of sending or receiving that much text.

You will also often see the term context window. It is how much text a model can lay out and look at at once. It is like a desk: the bigger the desk, the more papers you can spread out on it. To paste in a whole long document and ask about it, this value needs to be large.

2. The 12 services at a glance

• Copilot picks among several models depending on the kind of task, so its flagship model cell is empty.
• Seed 2.1 in the Doubao row is the flagship model announced by ByteDance Seed.
• Each service is covered to a different extent. For Meta AI, only the model and open weights are covered. For ERNIE and Doubao, pricing is not covered. This is because only values written on official pages were copied.
• Perplexity, a search-style service, is left out because its official pricing page and policy could not be checked.

3. Free and paid pricing

3-1. Personal plans

Below are the five services whose official pricing pages list values.

• Google AI Pro appears on the plans page with only its name and features, and no won price. Check the price on the checkout screen.
• Grok also has SuperGrok Lite, SuperGrok Heavy, and Business plans. These three appear on the pricing page table by name only, so check their prices on the sign-up screen.
• Copilot's paid features come with a Microsoft 365 subscription. You use Copilot inside Word, Excel, and PowerPoint. Storage is 1 TB for Personal and 2 TB for Family and Premium.
• Dollar prices are from the US pages. If you pay from Korea, the amount may differ because of taxes and exchange rates.

What to know when you start free: every free plan limits how many times you can use it, and answers slow down at busy times. Paying mostly buys you "more use, better models, faster." Try two or three services for free first, then pay for only the one you reach for most.

3-2. API pricing

This is the cost of calling a model from a program. Prices are per 1 million tokens, and the text you send (input) and the text you get back (output) are priced differently.

• DeepSeek's values are peak-time prices. Off-peak, they are half.
• Gemini 3.1 Pro and Grok 4.7 cost more when one request sends more than 200,000 tokens.
• Some models can be called for free. Gemini 3.8 Flash has a free tier on the API, and Z.ai offers GLM-4.7-Flash, GLM-4.5-Flash, and GLM-4.6V-Flash for free. What you send on Gemini's free tier is used to improve products.

Figure 2. The values in the table above as bars, ordered from the most expensive input. Dark bars are input, light bars are output (diagram, labels in Korean)

Prices split by model tier more than by country. The lowest rows are GPT-6 Luna and Claude Haiku 5.5 ($0.1 input), and the highest is GPT-6 Astra ($10 input). Among models in the $1 to $4 input range, DeepSeek-V4-Pro ($1.32) and GLM-5.3 ($1.4) are on the low side. A cheaper model is not guaranteed to suit your work better, so send the same question to two models and compare the answers before you decide.

4. What each one is good at

Below is what each company itself highlights on its official pages. It is not a user review or a ranking. Benchmark scores change with how and when they are measured, so they are not included in this guide.

Storage locations for both the US and Chinese services are collected in the tables in section 5. Grouped again by type of work, it looks like this.

• Writing and long documents: models with a large context window have the edge. In the table in section 3-2, every model except Grok 4.7 (500,000 tokens) is around 1 million tokens. In the ChatGPT app, how much you can send depends on the plan. With the reasoning models, Plus takes about 320 pages and Pro about 680 pages.
• Coding: a coding tool is easier than pasting code into a chat window. A coding tool reads and edits the folders on your computer directly. Claude Code comes with Claude Pro and up, and Codex is included in a limited form even in ChatGPT's free plan. Google has a tool called Antigravity, SpaceXAI has Grok Build, and DeepSeek has Harness.
• Search: check whether answers come with source links. Grok searches the web and X in real time, and Claude, ChatGPT, Gemini, and Copilot also search the web.
• Images and video: ChatGPT (GPT-Image-2.5), Gemini (Nano Banana, Veo 3.1), and Grok (Imagine) create them inside the chat window. Claude only reads images.

5. Privacy: where your data is stored and whether it is used for training

Text you type into an AI chat window goes to that company's servers. It is a bit like mailing a postcard. Who receives it, how long they keep it, and what they use it for are written in the terms. There are two things to look at: whether your chats are used to train models and which country's servers store them.

Figure 3. The path text takes after you type it into a chat window, and three ways to cut that path (diagram, labels in Korean)

The two tables below copy the same five items from the privacy policy of each personal app. Where a policy does not cover an item, the cell has a dash (—).

5-1. US services

• Grok's privacy policy does not say which country stores the data. For ChatGPT's storage location, check the OpenAI privacy policy.
• The Copilot row is based on the new app released in August 2026. In the old app, chats may be used for training, and you turn that off in settings.
• Claude deletes chats you delete from its servers within 30 days, and Gemini's activity history is deleted automatically after 18 months by default.
• Meta AI and ERNIE are not in either table because their app privacy policies are not covered.

Business plans and APIs follow different rules. OpenAI's Team, Enterprise, and API, and Claude's Team plan, do not use what you send for training by default. If you handle company data, read the terms of the business plan, not the personal plan.

5-2. Chinese services

• The Qwen row describes the API policy. The chat app has its own policy, so read the app's policy when you sign up.
• DeepSeek has a separate privacy policy in Korean and cites articles of Korea's Personal Information Protection Act.
• Kimi's and Doubao's policies are documents in Chinese.

Here is what to keep in mind when reading the two tables.

• The law that applies depends on the storage location. If the servers are in China, Chinese laws and regulations apply to that data. If they are in the US, US law applies. This does not mean one side is better. It means you should know which country's rules your data falls under when you choose. For someone living in Korea, both US and Chinese servers are abroad.
• Even one company can differ by channel. Some, like Z.ai, have a Singapore company run the channel for users outside the country. Check the operating company name and the storage location sentence at the top of the policy.
• Turning it off may not be a switch. Kimi and Z.ai need a separate request instead of a settings switch.
• Check your company's or school's rules first. Some workplaces and schools block certain AI services. Check the internal rules before you put work data in.

5-3. Rules for any service

1. Do not enter national ID numbers, passwords, card numbers, or the original text of contracts.
2. On the day you sign up, find the training-use item in settings and set it the way you want.
3. For something you want to ask only once and not keep, use a temporary or private chat.
4. Do not upload files that contain other people's personal information.

6. Open-weight models

You can download an open-weight model and run it on your own computer. Your text is not sent over the internet, so the worries in section 5 shrink. In exchange, you need a powerful graphics card, and you have to install and maintain it yourself.

• Anthropic's official model documentation lists no downloadable Claude model. Claude is used only through the apps, the API, and the clouds of Amazon, Google, and Microsoft.
• Even when you can download a model, the license decides how you may use it. Licenses differ by model, so if you want to use one to make money, read the license on the model page. In this table, the license name was checked only for OpenAI's gpt-oss.

If you want to use a Chinese company's model but would rather not send your data to servers abroad, you can run an open-weight model on your own computer. In that case, your text does not go to that company's servers.

7. Four ways to use them

The web is enough at first. When you feel "I repeat the same task on the web every day," that is the time to move to a coding tool or the API.

8. Korean language support

Most current models can take questions and answer in Korean. How natural they sound differs by service and cannot be judged from official materials alone. Send the same request for the kind of writing you do most (work emails, reports, blog posts) to two or three services and compare. Also check whether the menus appear in Korean, whether you can pay in won, and whether there is customer support in Korean.

9. A table for choosing

You do not have to pick one service and use it forever. Subscriptions can be canceled month to month, so you can try a different one each month.

10. Common mistakes and fixes

Summary in one line

Compare two or three services for free with the same question → check the training setting and the storage location → pay for only the one you reach for most → keep data that must not leave out of it, or use a model that runs on your own computer.

References

• Claude: Pricing
• Claude Docs: Models overview
• Claude Docs: Multilingual support
• Claude Help Center: What is the Pro plan?
• Anthropic Privacy Center: Is my data used for model training?
• Anthropic Privacy Center: How long do you store my data?
• Anthropic: Privacy policy
• ChatGPT: Pricing
• ChatGPT Docs: Pricing
• OpenAI API Docs: Models
• OpenAI: How your data is used to improve model performance
• OpenAI: Open models
• Google: Google AI plans (Korean page)
• Gemini API Docs: Models
• Gemini API Docs: Pricing
• Gemini Apps Privacy Hub (Korean page)
• Languages and countries where Gemini Apps are available (Korean page)
• Gemini API Docs: Gemini 3.8 Flash
• Gemini API Docs: Gemini 3.1 Pro Preview
• Google Privacy Policy (Korean page)
• Google AI for Developers: Gemma
• About Grok
• Grok: Pricing
• x.ai Docs: Models
• x.ai: Consumer FAQs
• x.ai: Terms of Service
• x.ai: Privacy Policy
• Meta AI for Developers
• Hugging Face: meta-llama
• Microsoft 365: Compare plans
• Microsoft Copilot
• Microsoft Support: Microsoft Copilot for individuals, your data, privacy, and responsible AI
• Microsoft Support: Microsoft Copilot for individuals, your activity history
• Microsoft Support: Microsoft Copilot for individuals, your privacy controls and choices
• Microsoft Privacy Statement
• DeepSeek
• DeepSeek API Docs: Models & Pricing
• DeepSeek Privacy Policy
• Hugging Face: deepseek-ai
• Alibaba Cloud Model Studio: Models
• Alibaba Cloud Model Studio: Privacy notice
• Hugging Face: Qwen
• Kimi
• Kimi API Docs: Pricing
• Kimi Privacy Policy (Chinese)
• Hugging Face: moonshotai
• Z.ai Docs: Pricing
• Z.ai: Privacy Policy
• Hugging Face: zai-org
• Z.ai (Zhipu AI) official site
• Baidu ERNIE Blog
• Hugging Face: baidu
• Doubao Privacy Policy (Chinese)
• ByteDance Seed (Chinese)