There are two main ways to share ChatGPT related services with friends: one is to create ChatGPT API or Azure OpenAI Service by yourself, and use clients such as OpenCat or ChatGPT Next Web to allow friends to access directly. The other is to recommend services such as Poe or OpenRouter.AI to let friends register and pay for their own use.
1.1 How to choose a model provider?
OpenAI ChatGPT API
OpenAI ChatGPT has four main versions:
- CChatGPT free version, which can only use GPT 3.5;
- ChatGPT Plus, $20/month, can be subscribed through App Store or Google Play in-app purchase; after subscription, GPT4 messages are limited to 40 messages every 3 hours;
- ChatGPT Team, $30/month/user, $25/month/user for annual payment; payment must be made by credit card;
- ChatGPT API, payment is based on usage, payment must be made by credit card; plug-ins cannot be used, and there is no ability to access the Internet; ChatGPT API is not open in mainland China, Hong Kong and other regions, so a credit card or savings card issued abroad is required to successfully deduct money. Generally, there are several solutions:
- Overseas credit cards/savings cards; Singapore OCBC Bank supports online account application with passport in the Mainland. After the account is opened, you can apply for a debit card. For specific capital costs and time consumption, please refer to the online tutorial;
- Virtual card platforms such as NOBEPAY can generate virtual cards by recharging RMB after real-name authentication; the exchange rate is very poor and the handling fee is high, but the overall cost of using the card is controllable;
- OneKey Card, Dupay and other virtual debit card service providers in the cryptocurrency field; recharge USDT and other cryptocurrencies in your account to open a virtual dollar card for use; on the one hand, these channels are costly; I have an OCBC bank account in Singapore, but because I opened a 360 account on the web, the card was not sent by default, and I had to apply for a savings card later, so I didn’t do it because I thought it was troublesome. NOBEPAY virtual cards are valid for a very short period of time, only more than a year; at the same time, I don’t hold virtual currency, so it’s really difficult to deposit money into OneKey Card. On the other hand, OpenAI has stricter risk control, and virtual cards are not once and for all after binding, and it is very likely that the deduction will fail next month. After recharging twice, I used it briefly and didn’t continue to use it.
Azure OpenAI Service
Product introduction page | Application form page Azure OpenAI Service is an OpenAI GPT service provided by Microsoft Cloud, serving enterprise users. The model release is generally half a year to a year later than OpenAI’s own service, and there are more security reviews based on OpenAI. The advantage is that deployment and payment are convenient, and domestic Visa/Mastercard can be used for payment.
Azure OpenAI Service is not yet open to individual developers. You need to apply for it in the name of an enterprise user using your own domain email address. You can submit the application on the application form page. Generally, you will be notified of the successful application within a few days. There are many application tutorials on the Internet, so I will not repeat them here.
After successfully applying for Azure OpenAI Service, you can use GPT4 directly. If the resource was created before GPT4 was opened, it will not be updated by default, and a new resource can be created. According to my experience, the following two points need to be noted:
- Not all regions of Azure OpenAI Service support GPT4. For specific regions, please check the official help page. The region here refers to the region where the resources are located, and has nothing to do with the region where the account is located. Enterprise accounts in China can use all OpenAI Service regions.
- Some early versions of GPT4, such as “gpt-4 (0613)”, have been discontinued, but they still appear in the options when creating a model. After selecting them, an error message will be displayed indicating insufficient resource quota. Just be careful to avoid them. The official help page will also list discontinued early versions of GPT4.
Google Gemini
Product introduction page | Application page Gemini API was released at the end of 2023. There is still a free quota available. There are 60 requests per minute for free trial, which is enough for daily sharing with friends. Most of the clients introduced below are already supported.
In fact, Google Bard is still free, and the effect of searching web pages and summarizing content is better than ChatGPT. You only need to use scientific Internet access to access it. The author often uses Google Bard to summarize web page data when writing articles.
Anthropic Claude
Product introduction page | Price page Claude 2.0 is comparable to GPT 4.0; Claude 2.1 is comparable to GPT-4 Turbo; Claude has more advantages in speed and price, but Claude 2.0 is more suitable for simple instructions. Because there is less research on each model and Claude is almost not used, I will not introduce it here. The advantage of Claude is that it is more affordable and more convenient to pay.
1.2 Which third-party client should I choose?
The artificial intelligence services introduced above basically require scientific Internet access to access. When choosing a client, you need to consider whether it supports proxy forwarding services, so that end users can access them directly after configuration.
OpenCat
Supported models include:
- OpenAI ChatGPT API (directly supported and can create Team version private deployment)
- Azure OpenAI API (directly supported)
- Anthropic Claude API (directly supported)
- Google Gemini API (direct support) ServerCat is a third-party AI Chat client developed by the well-known Twitter user Baye, supporting iOS and macOS. There are two payment methods, OpenCat Pro unlocks enhanced features permanently for $9.9; OpenCat Cloud is subscribed monthly, with limited use of ChatGPT 4. After unlocking OpenCat Pro with an in-app purchase of $9.9, you can create a Team deployment private version, fill in the ChatGPT API you created and transfer it through a private server, and then share it directly with other friends for use. The shared users do not need to use a proxy when using it. OpenCat should have no obvious shortcomings except that it does not support Windows and Android. I has been sharing ChatGPT with friends through the ServerCat Team version. However, after it was re-listed at the end of 2023, the client downloaded from the app store in mainland China was restricted and could not use the models released by foreign companies, so it was abandoned afterwards.
NextChat ( formerly ChatGPT Next Web)
Supported models:
- OpenAI ChatGPT API (direct support)
- Azure OpenAI API (direct support)
- Anthropic Claude API (direct support)
- Google Gemini API (direct support) ChatGPT Next Web is the most starred and discussed project among many third-party open source AI Chat clients. Because it supports free one-click deployment with Vercel, it can be deployed without VPS and domain name. The free access quota provided by Vercel is enough for small-scale use. I currently uses Vercel to deploy ChatGPT Next Web and share it with friends. Because the payment of ChatGPT API is too troublesome, ChatGPT 4 is disabled. At present, the project officially supports Azure OpenAI and can be configured by yourself.
chatbox
Chatbox is similar to NextChat, but the difference is that Chatbox was originally focused on the client, so it supports both desktop and mobile terminals, and it supports more models than Chatbox. The advantage of NextChat is that you can directly share the generated code with your friends after deployment, and then develop the client. If you use chatbox, you need to use Cloudflare Workers or other services for proxy to facilitate user access.
Chatbot-ui
Similar to NextChat, Chatbot-ui is also a well-known third-party open source AI Chat client. It is undergoing a major revision recently, so the client and web version may change significantly. I will wait until the new version is stable before introducing it.
LibreChat
LibreChat is also a well-known third-party open source AI Chat client. It supports one-step deployment through railway. At the same time, the function updates quickly, and it already supports the functions of uploading and analyzing pictures of GPT-4 and Gemini Vision, as well as functions such as accessing the Internet and plug-ins.
1.3. How are APIs managed?
Both ChatGPT API and Azure OpenAI Service support recording usage and fees based on API or instance. However, it is also necessary to manage API and access passwords through a third party.
Cloudflare AI Gateway
Cloudflare integrates Workers and other functions. Its AI Gateway can directly manage the APIs of multiple platform models, cache responses, limit and retry requests, and provide analysis functions to help monitor and track usage. Cloudflare AI Gateway will be more inclined to commercial deployment in the future, but the current functions are relatively simple and can be tried.
2. Paid third-party clients/services
If your friends use ChatGPT a lot and have difficulty accessing the proxy, you can recommend some mature third-party paid products.
MorseGPT
- Client: Web page, API
- Price: Charged by usage after recharging
MorseGPT is a simple wrapper service for OpenAI and other large models, supporting Alipay and WeChat purchases by volume. MorseGPT provides APIs for the OpenAI protocol, and can be used with foreign clients such as Chatbot-uiand TypingMind .
Poe
- Client: Web, Windows, macOS, iOS, Android are all supported
- Price: $19.99/month, often with special offers
- Supported models: OpenAI ChatGPT4; Anthropic Claude2; DALL-E-3, etc.
Poe is a well-known third-party AI Chat client that provides access to OpenAI GPT-4 and Anthropic Claude-2. It is popular in areas where OpenAI does not provide services, such as Hong Kong. After the paid package was released at the end of 2023, the traffic has declined slightly, and the current monthly visits are about 50 million / month. The price is similar to OpenAI’s subscription. Because the number of Chats provided in the paid package has been changing, the difference in package content has not been compared.
OpenRouter.AI
- Client: Web page, API
- Price: Charged by usage after recharge
OpenRouter.AI provides hundreds of language models. Its main business model is to allow users to choose the cheapest and most effective language model according to different usage scenarios. After recharging, users are charged according to usage. This model is becoming more and more popular, with monthly visits approaching 900,000. At the same time, OpenRouter also supports API access, and foreign clients such as Chatbot-ui and TypingMindalso support OpenRouter.AI’s API.
TypingMind
- Client: Web page, macOS client
- Price: Buyout, please check the official website for comparison of different packages
A private deployment version is also available. TypingMind is a third-party client with free basic functions and permanent enhanced functions for a one-time payment. It supports OpenAI ChatGPT API, Anthropic Claude API, and custom models. It should be noted that TypingMind is only a third-party client and cannot use GPT4 or similar models by itself. It needs to fill in the API of other platforms. TypingMind paid enhancements are available as plug-ins and web browsing.