1 Introduction

It's been a while since I've updated the AI learning map on my blog. It's not that my interest in AI has waned; quite the opposite. Many ideas and explorations have already taken shape in my mind. The real reason I stopped was actually very simple:Many practical scenarios ultimately cannot avoid the threshold of "paid API".

Although there are many free API services on the market, they either have very low call counts, outdated models, cumbersome deployment, or even fragmented experiences, which really makes me lose the motivation to continue "tossing". As for "paying", I certainly don't reject it.Unfortunately, many service providers exclude domestic users like us.Especially when it comes to payment, there are various risk controls, restrictions, credit card verification failures, and even the inability to apply for an API key.

Recently I finally got WordPress multi-active deployment (see:Home Data Center Series WordPress Multi-Active Architecture (Simplified Version) Implementation Plan in Personal Blog), and with some free time, I decided to return to exploring AI. First, I wanted to solve the most basic and troublesome problem:How to use paid API services of mainstream foreign AI stably and with low barriers.

2 Why did I choose OpenRouter instead of OpenAI?

2.1 Why did I choose not to use the official OpenAI API?

The GPT series has a benchmark status in the field of AI, but as a domestic developer/user,It is difficult to use these models directly through the official OpenAI API, the core issue is actually two words:Payment (of course there are regional access restrictions, but this can be solved through science or magic, so it is not a core issue).

OpenAI currently doesn't officially support Alipay or WeChat Pay, and even domestically issued credit cards (including dual-currency cards) can't be successfully linked or used for payments. As a result, many domestic developers have been forced to resort to "roundabout solutions":

• Or try to get an overseas credit card (time-consuming, labor-intensive, high threshold, and many risk controls)
• Either rely on various "magic payment" channels (such as virtual cards, recharge codes, payment intermediaries, etc.)
• What’s worse is that some services that were once still available (such as WildCard) have now gradually become “cold”.

These methods seem to solve the immediate problem, but they are not long-term solutions:They are neither stable nor compliant, and cannot support the long-term use of a formal project..

2.2 OpenRouter: A Leader in Multi-Model Access Platforms

After some research, my conclusion is:Instead of sticking to OpenAI's official channels, it's better to choose a reliable third-party compatible platform.In this type of platform,OpenRouter is one of the most recommended choices at present.

OpenRouter.ai is a technology startup headquartered in New York, USA. Founded in 2023, it is dedicated to building a unified multi-model access platform. It does not develop large language models itself, but rather integrates official APIs from major model providers such as OpenAI, Anthropic, Google DeepMind, Meta, and Mistral, providing a "neutral middleware" access method for developers and application systems. With OpenRouter, users can flexibly switch between different models using a unified API syntax, lowering the barrier to entry and avoiding the complex process of directly working with multiple vendors.

Led by OpenSea co-founder Alex Atallah and engineer Louis Vichy, the company is backed by top Silicon Valley investors such as a16z (Andreessen Horowitz), Menlo Ventures, and Sequoia Capital, and completed a $40 million Series A funding round in mid-2024. The company employs a fully remote work model, with core members drawn from across the United States and Europe, maintaining the agile pace of a startup team.

At the product level, OpenRouter's strengths lie not only in its wide selection of models and low barrier to entry, but also in its exceptional developer ecosystem friendliness. In addition to standardized API access, it also offers comprehensive documentation, flexible request pricing, and real-time model routing. This is particularly important for developers looking to integrate AI capabilities into their applications, especially in China, where directly accessing official APIs can often be fraught with obstacles due to payment, network, and risk control issues.

It is worth mentioning that OpenRouter does not "proxy" model data, but strictly completes request reasoning through the official authorized API. The data will not be intercepted or stored by the platform. It is only a "distributor" of model requests and does not participate in the underlying calculations or response modifications. This model also makes it more convincing in terms of stability and security. In other words,The GPT, Claude, Gemini and other models you call are actually still provided with inference services by their respective official systems. OpenRouter is just the "routing layer" in the middle rather than the "camouflage layer"Compared to some "informal channels" that rely on reverse engineering or crawler simulation, OpenRouter is closer to an "official transfer" mechanism. There is no need to worry about common risks such as data leakage, large model delays, and interface failures. It is also more suitable for production deployment in formal scenarios.

Precisely because of this positioning, OpenRouter has rapidly attracted millions of developers over the past year, becoming one of the most active and mature AI multi-model access platforms. For domestic users, its key appeal lies in its compatibility with dual-currency cards issued by domestic banks (such as Visa and Mastercard) and support for Alipay and WeChat top-ups. Furthermore, it supports direct access to domestic networks (though this is not recommended, for reasons discussed later), virtually eliminating the "non-technical barriers" that often plague traditional API access.

---

OpenRouter is not the only multi-model access platform on the market that supports Alipay/WeChat and supports direct domestic connections. For example, OhMyGPT, which is familiar to domestic users (and is also a third-party multi-model API provider I previously recommended), can be referenced in my previous article for those who are not familiar with OhMyGPT:Home Data Center Series Starts the AI Journey: A Detailed Introduction to Local Big Language Model UI and Big Language Model API Providers), which also supports Alipay/WeChat payment and direct connection to the domestic network - but I still recommend OpenRouter for 4 reasons.

1,CredibilityOpenRouter is backed by renowned entrepreneurs and top Silicon Valley investment institutions, and has established long-term partnerships based on official APIs with companies like OpenAI, Anthropic, and Google DeepMind. This means its API call process is transparent, data cannot be privately intercepted, and service stability is further guaranteed.

2,Industry recognitionIt is not only highly regarded by many developer communities, but also natively supported by a large number of third-party AI tools, such as the well-known WordPress plugin AI Engine It has built-in OpenRouter access options, which is rare among similar platforms:

This ecological advantage allows you to switch between different applications more easily without having to maintain multiple sets of configurations yourself. This is why I eventually switched from OhMyGPT to OpenRouter.

3.The balance between globalization and localizationIt maintains the standards and security of an international platform while optimizing the payment and network experience for domestic users. Compared to some "gray solutions" that rely on reverse engineering, it is more sustainable and will not fail or limit functionality at any time.

4.Transparent pricing and low premiumsOpenRouter's token pricing is essentially the same as OpenAI's official pricing, with only a 5.5% handling fee charged for deposits. Platforms like OhMyGPT typically add approximately 10% to the official price as a cost and profit. For high-frequency calls or large-scale model scenarios, this difference can accumulate to a significant level.

All in all, OpenRouter is a "easy and secure" choice for me, which is why I finally decided on it as the recommended multi-model access platform after research.

---

2.3 Comparison of Usability of Mainstream AI API Platforms for Domestic Users (2025 Edition)

To more intuitively illustrate the advantages of OpenRouter, I compiled a comparison table of common AI API service providers, including both multi-model access platforms and official model APIs. This table covers several key factors, including registration difficulty, connectivity, payment methods, and recommendation level:

name	Registration available	Direct connection available in China	Supported payment methods	Recommendation level	Remark
OpenRouter	✅	✅	Dual currency credit card (recommended), Alipay/WeChat	⭐⭐⭐⭐⭐	Easy to access, supports multiple models, has high ecological recognition, and is the first choice for domestic developers.
API2D	✅	✅	Chinese bank cards, Alipay, WeChat and other third-party channels, Stripe	⭐⭐⭐⭐	Close to native OpenAI interface, slightly more expensive than OpenRouter
OhMyGPT	✅	✅	Alipay, WeChat, domestic bank cards	⭐⭐⭐⭐	Targeting domestic users, with transparent pricing, but the number of models and update speed are inferior to OpenRouter, and its ecological influence is weaker.
Hugging Face	✅	✅	Overseas foreign currency credit cards	⭐⭐	The free quota is limited, and the model needs to be deployed by yourself. It is suitable for users who are familiar with the deployment process.
Claude (native)	❌	❌	Only overseas credit cards and addresses are accepted (difficult to register domestically)	⭐⭐	Indirect access via OpenRouter is recommended
Gemini (Google)	❌	❌	Overseas Google Pay requires an overseas credit card	⭐⭐	The official interface is fragmented and the experience is inconsistent. It is basically impossible to connect directly in China.
Azure OpenAI	❌	✅	Binding foreign currency cards and corporate invoices to corporate accounts	❌	Enterprise-oriented, cumbersome access, and requires approval process
OpenAI official	✅	❌	Foreign currency credit cards issued overseas only(Credit cards issued by domestic banks are basically unavailable)	❌	It is difficult to register and bind a card, and there are many API restrictions. Direct use is not recommended.

In addition to the GPT series, it also integrates popular models such as Claude, Gemini, Mistral, Command R, and LLaMA.Unified interface standards, pay-as-you-go, complete documentation, no magic payment required, which is very suitable for domestic users to quickly access, develop and test, or integrate products.

More importantly, it not only supports direct domestic access, but is also extremely domestically friendly in terms of payment methods:

• ✅ Support dual currency credit cards (Visa/Master)
• ✅ Support Alipay, WeChat scan code recharge (Officially provided by OpenRouter, but not directly displayed on the checkout page. It needs to be triggered through a specific entry.This method has a stable process and low threshold, making it very suitable for domestic users. For details, please see the "Credits Page Detailed Explanation" section below.)

OpenRouter provides domestic users with a truly "out-of-the-box" experienceNo registration or card binding required, no risk control required, multiple payment models supported, pay-as-you-go, and recharge upon use. Compared to other payment solutions that repeatedly "twitch" payment methods, this is undoubtedly the most worry-free and most suitable solution for practical development.

3 OpenRouter Account Registration

OpenRouter account registration is very simple, on the official website (https://openrouter.ai/) You can register by yourself according to the following graphic process (currently it can be directly accessed in China):

I have a gmail account, so I just click on the google logo:



In this way, the openrouter account is registered successfully:

After that, you need to set various common parameters of the account.

---

Generally speaking, there is no problem in registering on the OpenRouter official website by directly using domestic broadband, and no special network environment is required. However, if you access the website in a "scientific Internet access" state, and the DNS settings are set to Fake-IP mode, and you are using the Chrome browser, you may suddenly encounter a cold prompt.Sorry, you have been blocked.":

This is very mysterious - it's not that the IP is blocked, nor is there a problem with the account, but that some underlying mechanisms combined together triggered Cloudflare's risk control strategy (OpenRouter's website is based on Cloudflare).

Simply put, this is because in Fake-ip mode, the target IP accessed by the browser is actually forged, and Chrome actively attempts to verify the connection. As a result, Cloudflare mistakenly judges it as a "suspicious request" or even "automated behavior" and directly blocks it.

The solution is not complicated, you can:

• Try switching to Firefox or Safari browser temporarily;
• Or, more thoroughly, set openrouter.ai to direct connection in Clash to avoid DNS interference;
• If you are using a Clash-based OpenWRT router, you can also set a bypass rule for the domain name.

This question is actually quite interesting, involving the working principle of Fake-IP, Chrome's DNS behavior, and Cloudflare's blocking judgment mechanism. I will write a special article later to analyze this phenomenon in detail.

---

4 Introduction to OpenRouter Account Settings

4.1 Overview

After registering an OpenRouter account, the next step is to make some necessary settings for the account. The backend settings panel of OpenRouter is not complicated and is divided into the following 7 parts:

• Settings: Basic settings page, where you can modify your nickname, avatar, default language model, etc., suitable for priority configuration after registration;
• Credits: A place to check balance and spending history, and also the top-up portal, including the function of purchasing points through domestic channels (such as Alipay);
• Presets: Preset configuration, supports defining custom parameters (temperature, context length, etc.) for different models, which can be easily referenced when calling the API;
• API Keys: Used to generate and manage API keys, which are the core credentials for calling OpenRouter;
• Provisioning Keys: Mainly used in team collaboration or SDK integration scenarios, it can provide fine-grained control over permissions;
• Integrations (BYOK): Supports "Bring Your Own Key" mode, suitable for users who already have a model API key to connect it to OpenRouter's unified interface;
• Privacy: Privacy options, users can choose whether to allow OpenRouter to record request logs and other behaviors.

These sections have different focuses in actual use: if you just want to call the model through the API, you only need to configure the API Keys to start using it; but if you want to manage different projects in detail, use Alipay to recharge, or use presets to manage multiple model parameters, then the other sections will also be very useful.

Starting from the next section, we will introduce these functional blocks in turn, starting with the basic and important account settings page:Settings.

4.2 Detailed explanation of the Settings page

---

✅ Account (Account Management)

This column can be clicked to manage your account information, such as:

• Change login email/password;
• Set up or change security verification methods (such as enabling 2FA);
• Completely log out and delete the account.

Although it is unlikely to be changed in daily use, it is recommended to set up security items at the beginning to prevent the account from being abused.

---

✅ Organization

OpenRouter supports creating "Organizations" to facilitate team collaboration, unified billing, and other scenarios. For individual users, you can ignore this option; however, if you plan to have multiple members share an API key or use unified billing in the future, you can create an organization and invite other accounts to join.

---

✅ Low Balance Notifications

When your account balance is low, the system will automatically send you an email reminder. By default, it will be sent to your registered email address. It is recommended to keep this function enabled to prevent unnoticed call failures.

---

✅ Allowed Providers (allowed model providers)

OpenRouter integrates with many model providers, such as OpenAI, Anthropic, Mistral, Cohere, etc. If you only want to use a few of them, you can manually select the allowed providers here.

There is also a “Always enforce” Options:

• After opening,Only the providers you check here will be used;
• When not enabled, all available providers that are not ignored are used by default.

This feature is suitable for people who want to precisely control the direction of the request (for example, they only want to use OpenAI and do not want to use any open source models).

---

✅ Ignored Providers

In contrast to the above, here you can select the model providers you do not want to use. For example, some people do not trust certain niche providers, or feel that a certain provider is too slow to respond, so you can exclude them here.

---

✅ Default Provider Sort (default provider sorting logic)

OpenRouter will automatically select a model provider based on certain rules when you do not specify which provider to use. This "default sorting logic" can be controlled here. The default values are:

• Default (balanced): Automatically balances lower prices and service stability, suitable for most users.

In addition to the default options, you can also choose the following sorting strategies according to your needs:

• Price (cheapest first): Give priority to the provider with the lowest price, suitable for cost-sensitive users;
• Throughput (highest first): Give priority to providers with the highest throughput (concurrent processing capability), which are suitable for application scenarios with a large number of concurrent requests;
• Latency (lowest first): Give priority to providers with the lowest response latency, which is suitable for real-time interaction scenarios with high response speed requirements.

Generally speaking, if you don't have a particular preference, just keep the default; but if you are pursuing the ultimate cost-effectiveness or are extremely sensitive to response speed, you can also flexibly switch the sorting logic according to your needs.

---

✅ Default Model

This is an important setting. You can specify a default model here. In the future, when you call the OpenRouter API, if you do not manually specify a model, the one set here will be automatically used.

ClickClick here“You can open the model list and see which models are supported and what the price is for each model.

⚠️ Tips:

• If you are using some applications that are very sensitive to model selection (such as automatic summarization, translation robots, etc.), it is recommended to explicitly set the default model;
• If you frequently switch models, you can also leave it unset and specify it manually each time you request it.

---

summary

Although the Settings page is generally "basic configuration", there are several key options that will actually have a profound impact on the behavior of OpenRouter when you call it later, especially the "Default Model" and "Provider Selection" sections.

It is recommended to go here after registering an account:

• Set up your most commonly used default model;
• Allow or ignore specific providers based on usage preferences;
• Enable balance alerts to avoid API request failures.

In the next section we will go into Credits page, learn how to check your balance, top up, and make payments through domestic channels such as Alipay.

---

4.3 Credits Page Detailed Explanation

4.3.1 Overview

Credits pageIt is the main entrance for OpenRouter users to manage account balance, recharge and consumption. In simple terms, it is the main entrance for OpenRouter users to manage account balance, recharge and consumption.Wallet Center, almost all billing-related operations can be completed here:

The overall interface is very simple and intuitive, and mainly includes the following parts:

• Account Balance (Available Credits): Displays your current available balance in USD. Fees incurred when calling models will be deducted from this balance in real time.
• Recharge entrance (Buy Credits): Click the "Add Credits" button to jump to the recharge interface, which supports foreign currency credit cards, some RMB cards, and recharge via Alipay/WeChat redemption codes (detailed introduction in Section 5.3.2).
• Auto Top-Up: Allows you to set an automatic recharge rule to automatically deduct renewal fees when the balance falls below the specified threshold, avoiding interruptions to API calls due to depleted balance (explained in detail in Section 5.3.3).
• Payment History: Displays all your past recharge records, including time, amount and payment method, for easy checking.
• Consumption Details (View Usage): You can view the billing details of each API request by model, time, project, and other dimensions to help you analyze and control costs.

Overall,The Credits page has clear functions and simple operationEven first-time OpenRouter users can quickly get started. You can think of it as a "transparent billing dashboard" that provides a complete view of the flow of funds for your model calls.

4.3.2 Buy Credits (Recharge Portal)

If you do not want to enable the auto-recharge feature,Buy Credits This is your main entry point - used to manually top up your account balance and flexibly control capital investment as needed.

OpenRouter provides a variety of recharge methods, covering traditional payment channels and cryptocurrencies, suitable for developers in different regions and with different preferences.

4.3.2.1 Traditional Payment Methods (Recommended)

ClickAdd Credits" button, you will see the following four mainstream recharge methods:

1. Card
   
   

The most commonly used and recommended method, supports binding Visa / MasterCardDual Currency Credit CardMost overseas credit cards can be used to complete the payment, including those issued by some domestic banks.Dual Currency Card(Whether it is available depends on the risk control strategy of the issuing bank).

Note:RMB single-currency cards are almost incapable of completing paymentsTherefore, if domestic users choose this method, it is strongly recommended to use aCredit cards with foreign currency payment capabilities.

2. Bank
   
   
   
   Payment via US local bank account (ACH), suitable for US users,Not suitable for domestic users.

3. Cash App Pay
   

   

Payment apps for US users need to be linked to US mobile phone numbers and bank accounts.Not available for domestic users.

4. Amazon Pay
   
   

Need to bind Amazon account and international credit card, the function coverage is limited,The success rate of domestic use is low.

Overall,For domestic users, the only practical way is Card (foreign currency credit card), it is recommended to give priority to this path for binding and payment.

4.3.2.2 Cryptocurrency Payment (Use crypto)

If you don't have access to a credit card, you can also choose toCryptocurrencyTo recharge your OpenRouter account. On the recharge page, clickUse crypto" Switch to open the encrypted payment channel, including BTC, ETH, USDT Wait for mainstream crypto assets to complete the transfer:

Unlike traditional payment methods, OpenRouter's cryptocurrency payment process is supported by third-party platforms (such as Coinbase), which provides a smooth overall experience and flexible access methods. However, it should be noted that this method willAn additional handling fee of 5% will be chargedThe official description states that the actual payment amount is calculated as follows:

total_cost = credits_requested / (1 - 0.05)

That is, if you want to top up your account with $100, you will actually need to pay about $105 in cryptocurrency.

After entering the payment process, OpenRouter provides a variety of wallet options for users to choose from, including:

• Coinbase: Suitable for users who already have a Coinbase account, you can directly jump to pay;
• Coinbase Wallet: Used to connect to Coinbase's own crypto wallet;
• Smart Wallet: A built-in wallet that does not rely on browser plug-ins, suitable for users who are not familiar with Web3 operations;
• MetaMask: The most common browser extension wallet, recommended for most desktop users.

In addition, you can click "Connect another wallet" to access more types of wallets:

Support includes Rainbow, WalletConnect, Browser Wallet, Argent, Ledger (hardware wallet), Phantom, Rabby, Trust Wallet, Uniswap Wallet, and many other mainstream wallets. Whether you hold coins on Ethereum, Polygon, Solana, or other mainstream chains, you can basically find a suitable payment path.

If you are already familiar with Web3 wallet operations, or already hold stablecoin assets (such as USDT/USDC), encrypted payment is undoubtedly a safe, fast and real-name verification-free way to pay.However, for most domestic developers, if they have no experience using cryptocurrencies, it is still recommended to use dual-currency credit cards first, which is simpler to operate and has lower risks.

4.3.2.3 Alipay/WeChat QR code scanning (transfer purchase)

In addition to the officially supported card payment and cryptocurrency methods, OpenRouter also provides aA relatively hidden but very user-friendly "third-party payment" channelIn this way, you can use Alipay/WeChat scan code paymentComplete the recharge directly.

This functionHidden by default, you need to follow the path below on the Credits page:

1. OpenOpenRouter's Credits page;
2. Click the Add Credits button.
3. At the bottom of the pop-up top-up window, click Use one-time payment methods
   
   
4. A new area will open on the page, providing the option to scan and pay with Alipay/WeChat
   
   
5. Follow the instructions and go to the external payment platform (Alipay or WeChat) to complete the purchase;
   
   
   
   
   

   ⚠️ Notes

• The payment process will jump to the OpenRouter accessThird-party cooperation platform, such as the payment page for Alipay/WeChat scan code; this entry is currently accessible through the official UI page (Add Credits → Use one-time payment methods), so it can be regarded as Platform-approved transfer methods;
• After payment is completed, the recharge amount will beDirect depositYour OpenRouter account balance, usually delayed by no more than 1 minute;
• Service fees are included in the total price(For example, if you pay 78 RMB and receive 10 USD, the implicit exchange rate is about 7.8). No additional fees will be charged, but the exchange rate will be different from the mid-price.
• Although it is not the native payment channel of OpenRouter, since the entrance is from the official website interface,This method has been widely verified in the community as stable and secure., suitable for novices or small-amount recharge users who do not have foreign currency credit cards.

Recommended group

• Foreign currency credit cards cannot be used;
• Unfamiliarity or inconvenience with using cryptocurrencies;
• I hope to complete the top-up through familiar domestic payment methods (Alipay/WeChat).

In general, this transfer methodSimple operation, fast payment, no need to register an additional wallet, which is one of the more practical solutions for domestic users. However, since it involves third-party payment, please always remain vigilant and avoid paying through non-official portals.

4.3.2.4 Summary

Overall, although OpenRouter lists multiple recharge channels on the interface, there are actually only three options that are truly suitable for domestic users, and their applicable scenarios are different.

First priority: foreign currency credit card

If you have a credit card that can be used for foreign currency payments (such as a dual-currency card, multi-currency card, or full-currency card from Visa or MasterCard),It is strongly recommended to use Card to top up your account first.This is the most official, direct, and cost-effective recharge channel:

• The recharge process is convenient and the funds are credited instantly;
• No need for intermediaries or third-party tools;
• No additional handling fees, which is also the only “0 cost” recharge method;
• Can be bound to automatic renewal, API risk control friendly.

Second priority: Alipay/WeChat scan code recharge (transfer channel)

If you don't have a foreign currency credit card but still want to top up in RMB, you can also use Pay by scanning the QR code with Alipay or WeChatThis method is implemented through the transit channel embedded in the OpenRouter official page.The entry is located below the "Use one-time payment methods" button on the Add Credits screen.The operation process is simple and the experience is good.

Although it is a transit payment channel for platform cooperation (not a native official settlement channel), since the entrance is from the official website interface itself,Can be regarded as a controlled and relatively safe officially recommended methodAfter the payment is completed, the system will automatically recharge the balance to the account.No need to manually paste the recharge code, suitable for novice users who do not have credit card conditions, especially for small-amount experience scenarios.

• No credit card or cryptocurrency required;
• No handling fees, fast arrival (usually within a few minutes);
• Suitable for users who do not have the ability to pay in foreign currency.

Priority 3: Use crypto

For users who have a certain amount of crypto assets and are familiar with Web3 wallet operations, they can also choose to recharge through USDT, ETH, BTC and other currencies. This method has a low threshold, but requires5% platform fees, the arrival speed is usually faster. Suitable for technical users or special needs scenarios.

• No real-name or traditional banking system required;
• The handling fee is high, so frequent use of small amounts is not recommended;
• Supports multiple mainstream wallets, including MetaMask, WalletConnect, Coinbase, etc.

As for the ones listed in the interface Bank, Amazon Pay, Cash App Pay, etc.,Domestic users are basically unable to useMost of these options require a US local account or payment environment and are not universally applicable, so you can ignore them.

Conclusion: Foreign currency credit cards are preferred, scanning the recharge code is the most convenient alternative, encrypted payment is left to experienced users, and other methods can be ignored.

4.3.3 Auto Top-Up

In addition to manual on-demand recharge, OpenRouter also provides a relatively worry-free option:Auto Top-UpThis feature can automatically deduct funds from your linked payment method when your account balance falls below a certain threshold, replenishing your credits in a timely manner and preventing you from interrupting the use of the model API due to depleted balance.

This feature is particularly suitable for developers with more stable usage needs, such as access to production environments, regularly running tasks, or scenarios with high call frequency. It can effectively reduce the risk of "service interruption due to forgetting to recharge."

It should be noted thatAuto Top-Up only supports payment methods with automatic deduction capabilities, such as credit cards. This feature cannot be enabled for one-time payment channels such as cryptocurrency or QR code payment.

You can directly see the entrance to Auto Top-Up on the Credits page:

The specific method of adding payment methods has been explained in the previous section and will not be repeated here.As long as you have linked a supported payment method, you can set the amount and trigger conditions for automatic recharge.

Note: While automatic top-up is convenient, it also means the platform can initiate deductions without your active action. Therefore, it's recommended to ensure your card is secure and has sufficient credit before enabling it to avoid unnecessary financial losses due to misuse or abnormal balances.

4.4 Presets page details

4.4.1 Overview

In daily OpenRouter use, we often want to set different preferences for different tasks, such as using a fixed model, adding custom system prompts, adjusting build parameters, and controlling budgets. To reduce the hassle of repeated configuration, OpenRouter provides a "Presets" feature that allows users to save these settings as a preset template and recall them at any time.

The Presets page is simple yet incredibly powerful. You can name each Preset, add a description, set the default model or provider, and control advanced parameters like routing policy, privacy options, and pricing caps. Presets also support setting specific generation parameters, such as temperature, maximum token count, and duplicate penalties, making them ideal for users with specific requirements for output style or quality.

In short, Presets are the "basic settings" you lay for how you use your model. Once configured, you can call them up with a single click, eliminating the need to re-enter them each time. For those with regular usage needs, this will greatly improve efficiency.

---

Important reminder:

Presets is a API layer functionality, it will only take effect if you pass in the Preset ID when calling the interface. If you use a third-party tool (such as some WordPress plugins, LobeChat, Chatbox, etc.), it does not provide a Preset parameter entry, so these presets will not be automatically applied when calling, but will be executed according to the parameters set in the tool (Generally, these tools set the specific models and other information they use internally, so the settings in Presets are invalid for these tools.Therefore, Presets are suitable for use when writing your own code, using clients that support custom API parameters, or debugging interfaces, and can significantly reduce the workload of repeated configuration.

---

4.4.2 Create a new preset

On the Presets page, clickCreate Preset" button to start creating a new preset:

Next, I will explain the function of each setting item one by one, and demonstrate how to fill it out based on the needs of the "Writing Assistant".

1. Basic Info: preset basic information

The main thing here is to name the preset and write a brief description to make it easier for you to distinguish and manage it.

• Name: Write a memorable name, such as Creative Writing Assistant, Creative Writer, which supports Chinese.
• Slug: A unique identifier generated by the system. By default, it is the same as "Name". However, if "Name" is in Chinese, it is recommended to manually change it to English here.
• Description: Briefly describe the purpose, such as: "Suitable for Chinese content creation, with smooth and natural output."
  
  

2. System Prompt: System preset prompt words

This is aGlobal directives that control model behaviorYou can think of it as "default context" or "role definition".

For the writing creative assistant, you can set the following:

This will significantly affect the language style and response angle of the model output.

3. Models: Specify supported models

If you only want this preset to work with specific models, like ChatGPT-4o or Claude 3 Opus, you can check them here.

For the Writing Creativity Assistant, you can choose from the following models:

• OpenAI: ChatGPT-4o: Well-balanced, suitable for creative tasks
• Anthropic: Claude 3 Opus: Excellent language organization and excellent Chinese performance
  
  

Of course, you can alsoDo not select any model, indicating that this Preset is applicable to all models.

4. Provider Routing: Model Provider Selection and Routing Strategy

This part belongs to PresetAdvanced Options, it is not enabled by default and needs to be checked"Include Provider Preferences" Only after clicking on the "OK" button can you see the detailed configuration items. You can think of it as "telling the system which model providers I trust and whether to allow automatic switching to backup service providers if there is a problem." Most users can skip this because OpenRouter's default policy is smart enough to make reasonable choices between quality, speed, and price.

However, if you have the following requirements, manual configuration is recommended:

• Trust only a few providers (e.g., OpenAI, Anthropic)
• Hope to eliminate unstable service providers
• Want to clearly control the calling priority order of service providers (whether it is effective needs to be tested)
• Sensitive to privacy and do not want data to be collected or used for training

The parameters are described as follows:

parameter	Meaning and recommended settings
allow_fallbacks	Whether to allow automatic switching to backup service providers. It is recommended to keep it checked. Otherwise, an error will be reported directly in case of failure.
sort	Alternative routing strategy, keep the default
data_collection	Whether to allow data to be recorded by the service provider for training: - allow (default) - deny (more private) Selecting deny may limit the types of models
only	Whitelist, only allow these service providers to provide services
ignore	Blacklist to exclude unwanted service providers
order	Specify the calling order—— ⚠️ The actual sorting may not work, be careful!

---

The order field should allow users to specify the order in which multiple model providers are called, such as which provider to prioritize and which to try next if a failure occurs. However, my testing has shown that this feature isn't behaving as expected.

Specifically: After selecting multiple service providers to add to the order list in the interface, the user cannot manually drag and adjust their order. The system seems toAlphabetical sortingThis means that even if you choose OpenAI first and then Anthropic, it is very likely that Anthropic will be called first - simply because it comes first in the alphabet.

This is not clearly explained in the official documentation, so I can't confirm whether this is a UI limitation, a design flaw, or just the way it's done.Out of caution, we should assume that the current order field cannot precisely control the order of calls.

Therefore, if you only want to use a specific provider, such as OpenAI, you should select only that provider in the "only" field to avoid mixing in other options. If you need to set up multiple backup providers, don't rely on the "order" field to strictly control the order of calls. A better approach is to make appropriate use of the "only" and "ignore" fields to narrow down your trusted providers and keep "allow_fallbacks" enabled so that if the preferred provider becomes unavailable, the system can automatically switch to other available providers to continue processing requests.

---

Correct setting method (recommended)

If you want to prioritize a particular service provider and automatically switch to another one if it fails, please do the following:

You only trust OpenAI and Anthropic, hope to use it first OpenAI.

✅ The correct settings are as follows:

• Check ✅ Include Provider Preferences
• Enable ✅ allow_fallbacks
• Set order = [OpenAI]
• Set only = [OpenAI, Anthropic]
• Leave it blank ignore

5. Parameters: Model generation parameters

These are the detailed options for controlling the model's behavior:

The meanings of the parameters in the above figure are explained as follows. You can set them according to your actual needs:

Parameter name	default value	Function Description
Temperature	1.00	Control outputRandomnessHigher values result in more creative content; lower values result in more deterministic content. The recommended range is 0.2–1.2.
Top P	1.00	Also known as the "core sampling" parameter. Controls the generated contentDiversity, the smaller the value, the more the output is concentrated on high probability words.
Top K	40	Limits the maximum number of "high probability words" considered during sampling. Smaller is more conservative, larger is more flexible.
Frequency Penalty	0.00	Penalizes words that appear frequently in the text to reduce repetition. Suitable for generating more varied text.
Presence Penalty	0.00	Penalizing words that have already appeared in the text encourages the model to explore new content and reduce repetition.
Repetition Penalty	1.00	Another way to control repetition. A value greater than 1 will suppress repetition, while a value less than 1 may encourage repetition.
Max Tokens	1024	Control the maximum number of tokens generated in this answer (1 token ≈ 1 English word, or 1 to 3 Chinese characters).
Seed	random	Used forGenerate reproducible answersThe same seed and prompt can generate the same content on some models (model support required).

After completing the above settings, click the blue button in the upper right cornerSave Preset” to save. Next, you can switch directly to this "Writing Creativity Assistant" preset during conversations and enjoy a more customized experience.

4.4.3 Summary

So far, we have roughly understood the basic setting method of the model parameters.

These parameters are mainly used toFine-tune the model's output style, which allows you to more finely control dimensions such as randomness, repetition, and diversity of generated content. For ordinary users, the default values are sufficient for achieving relatively ideal output in most scenarios, so there's no need to worry about these technical details from the outset. However, if you have specific requirements for the style, structure, or consistency of the generated content, such as a more stable writing style, less repetitive content, or if you want the output to be reproducible, then adjusting these parameters appropriately can be very helpful.

Note that not every model supports all parameters. For example, some lightweight models may not support the Seed parameter, meaning that even if you set a specific random seed, your results are not guaranteed to be reproducible. Similarly, different models may respond slightly differently to parameters like Temperature and Top P, so some settings that are effective on one model may have little or no effect on another.

Therefore, if you are not familiar with the functions of these parameters, it is recommended to keep the default settings. Slowly observing and fine-tuning from actual use will be safer and easier to master the rhythm than making a large number of changes at the beginning.

4.5 API Keys page details

4.5.1 Creating Your Own API Keys

Although the API Keys page is very simple in function, it is a very critical part of the entire system. Any program, plug-in, or tool that calls Preset or model services must rely on the API Key for identity authentication and call authorization.

The process of creating an API Key is simple:

ClickCreate API Key" button, a small panel will pop up:

You just need to fill in the following fields:

• Give this key a recognizable name, such as "Writing Creativity Assistant";
• If necessary, you can set a credit limit in US dollars. Leaving it blank means there is no limit.
• In the advanced settings, you can also decide whether to Use of BYOK (Bring Your Own Key) ModelCalculate into the quota limit of this key (by checking Include BYOK usage in limit).

After clicking Create, the system will generate a key string.Will only be displayed once, please save it properly in time.

---

In the advanced settings, the "Include BYOK usage in limit" option is used to:Whether to include model calls in BYOK (Bring Your Own Key) mode in the quota limit of this API Key.

In simple terms:

• Check: Whether you use the OpenRouter model or the model key you bind yourself,All will be included in the quota control of this API Key;
• Unchecked: Only models invoked through the OpenRouter platform will be counted towards the quota, and BYOK mode will not be counted.

If you want to more strictly control calling behavior (for example, external tools abusing your bound OpenAI Key), you can enable this restriction; otherwise, you can leave it unchecked by default.

---

4.5.2 API Key Usage and Recommendations

When integrating OpenRouter into other platforms or tools,API Key is your identity "pass"It is commonly used in WordPress plugins, locally running AI tools (such as Chatbox, AnythingLLM), Python or Node.js projects to call model interfaces, or some command-line tools and third-party platforms that do not support OAuth login.

An API Key is essentially an access token that represents your account's calling permissions and quota consumption.Be sure to keep it properly to avoid leakage.Once exposed, others can call the model in your name, consume your balance, and even cause the risk of information leakage.

In practice, properly configuring credit limits can effectively improve security, especially in development, testing, or team collaboration scenarios. For example, if you deploy a preset or script for a friend and want them to try it out, but are concerned about large charges due to misuse, you can create a separate API key specifically for this purpose and set a credit limit of $1-2. This will minimize losses even if an abnormal call occurs.

Similarly, if you need to provide the Key to an external plugin or tool configuration,It is also recommended not to use your primary key, but to create a new dedicated key with a "single purpose and clear quota".This makes it easier for you to monitor and manage it later, and you can disable it at any time if necessary.

4.6 Detailed Description of the Provisioning Keys Page

Provisioning Keys OpenRouter provides a simplified authorization method that allows you to grant temporary access to others without exposing your primary account information or API key. It's ideal for scenarios where login isn't required, such as quickly building an online demo, embedding an AI chat component in a webpage, or providing an anonymous trial interface.

Currently in the actual interface, after clicking "Create Provisioning Key":

The system only requires you to fill in a name (Name), and then a Key can be generated directly:


Once this key is generated, you cannot view the content again, so it is recommended that you copy it and save it properly as soon as possible.

Although the current interface does not expose advanced configuration items such as credit limit or Preset binding, in actual use,You can still use this type of key with a specific preset through code or access tools., realize the behavior preset of model call.

It is important to note thatThe Provisioning Key has call permissions by defaultOnce leaked, it can be used by anyone. Therefore, it is recommended that you use it with a low-limit preset or current limiting mechanism to avoid resource abuse.

4.7 Integrations (BYOK) Page Detailed Explanation

This page is used to configure BYOK (Bring Your Own Key) Model access - as the name suggests,Use the API Key you applied for on other platforms (such as OpenAI, Anthropic, Mistral, etc.) to connect to OpenRouter for unified management and call.

What are the benefits of doing this? Simply put, it allows you to use the unified interface of OpenRouter.You can call their own proxy service or your own official API quotaThis is useful in the following scenarios:

• You already have your own OpenAI account, but want to continue using OpenRouter's presets, routing policies, usage statistics, and other features;
• You want to go through official channels on some models and use the access provided by OpenRouter on some models;
• Your organization needs to manage keys in a unified manner and does not want each application to be scattered and directly connected to various platforms.

On this page, each supported model platform (such as OpenAI, Anthropic, Mistral, etc.) will be displayed as a separate module. Click the "Options" button:

You can paste your platform API Key to bind:

After the binding is successful, the system will automatically mark these models as BYOK Source. After binding, these keys will not be publicly displayed, but you can disconnect or replace them at any time. In the actual call process,Whether to use BYOK or the quota provided by OpenRouter is determined by the preset you create or the call request parameters..

If you are worried about the BYOK Key being abused, you can use the permission restrictions in the Preset and API Key to control whether certain access paths are allowed to use the BYOK model.

4.8 Privacy page details

OpenRouter Privacy page Provides a series of options that allow you to more fine-tune how your data is processed during use, especially whether to allow the model service provider to record, use, or even train the model based on your input data:

The settings of the entire page are roughly divided into Paid Models,Free Models,Chat History and Analytics Cookies Four major categories:

1. Paid Models

You can decide whether to allow the use of Paid model providers who may use your input for training without anonymity.

• Enable providers that may train on inputs

If enabled, the system will allow calls to service providers that may collect input data for model training. If you want to use privacy with lower risks, it is recommended to turn this option off.

• Enable input/output logging

Once enabled, OpenRouter will save the input you submit through the system and the output data generated by the model, and you will getPrice discount for all platform models using 1%As an incentive (recommended to turn this off if you don't want anything logged).

2. Free Models

For free model service providers, the system provides additional privacy reminders:

• Enable training and logging (chatroom and API)

When enabled, calls to those Free model providers that may log, retain, or use your input for trainingThese services sometimes even make the prompt content you submit public, even though you remain anonymous in the system (if you value data confidentiality, it is strongly recommended to turn this option off).

3. Chat History

This section describes how OpenRouter handles chat logs:

• Your chat history isSaved locally in the browser, not uploaded to the OpenRouter server;
• The system will upload and save LLM input and output content only when you enable the "input/output logging" function.

This means that unless you manually turn on recording, all conversations are technically "only on your device."

4. Analytics Cookies

• Enable analytics cookies

Once enabled, the system will use some analytical cookies to help optimize website performance and user experience (such as recording page visit behavior, etc.). These cookies are not related to model calls and do not involve any input or output data.

In general,OpenRouter provides clearer privacy controls than most platformsYou can flexibly configure this based on your actual usage scenario. If you care about whether your inputs may be used for training and whether they will be saved, it is recommended to disable all options related to logging and training, and only use model providers that clearly mark "zero data retention".

4.9 Activity

The Activity page is where OpenRouter records and displays API call history, serving as a "usage log center." Here, you can view detailed information about each model call in chronological order, including the model name called, the time of the request, the number of tokens consumed, the corresponding fee, and the call status (success, failure, or rejection). If further troubleshooting is required, you can also expand a single record to view more specific request and response details, which is very useful when debugging APIs or analyzing usage costs. For users who need to manage costs, maintain usage statistics, or track faults, the Activity page is an important reference point, helping you understand call behavior.

Go to API Keys, click on the 3 dots to the right of the API Key you want to view usage, and then select "View Activity" in the menu that appears:

You can see detailed information of requests in any time period, which is very friendly:

5. Demonstration of actual usage scenarios of OpenRouter API key

5.1 AI Engine plug-in usage scenarios

The following example demonstrates the setup process using OpenRouter's "GPT-5 mini" as the chatbot's large language model:

If the "Quick Test" test is successful, the following is displayed:

Then set the default environment. This step is not necessary, but you can set it if you want to be lazy:

Then enable the chatbot feature:

The default chatbot mode is "images". Change it to "chat" mode and enter the prompt word:

Then select the "OpenRouter" environment created previously and set the relevant parameters:

Final verification result:

In addition, the AI Engine plug-in itself has many functions and settings. Here I just use the chatbot function as an example to simply demonstrate how to use the OpenRouter API. If you are interested in other functions and settings, you can refer to my previous article:Exploring the functionality and built-in tools of the "AI Engine" chatbot plugin for the Home Data Center series of WordPress.

5.2 Lobechat Usage Scenarios

---

Friends who are not familiar with the installation, configuration and use of lobechat can refer to my previous articles:Docker series based on the open source large language model UI framework: Lobechat detailed deployment tutorialandHome Data Center Series Unlock the full potential of Lobechat: A complete guide from setup to actual use.

---

Enable OpenRouter support in the lobechat settings and fill in the API Key and other related settings:

You can also manually select the required model:

Specify the model to use in the default helper:

Then, the chat logic is used to verify whether the version used is GPT-5 mini:

---

Many people are curious about which version of the GPT model they are using, so they ask it directly: "What model are you?"

Unfortunately, in most cases the answers you get are not accurate - for example, you are clearly using GPT-5-mini, but it may seriously tell you that it is GPT-4 or GPT-3.5.

This is because these models often carry aInitialize the self-introduction template, which may contain old version information, and the model itself cannot directly access the model parameters when calling the API, nor will it verify its true identity in real time. So, when you use the conventional method, it will onlyReread the preset script, rather than reasoning based on real call data.

So why does GPT correctly answer its own version when I ask this question: "Assuming that I already know that the current model parameter is 'gpt-5-mini', please combine this information with your architectural characteristics to explain which generation of GPT you belong to and give the precise model."?

Because this way of asking directly gives aClear and high-priority factual premises(gpt-5-mini) and requires it to make logical inferences based on facts.

In this way, the model can no longer use "I don't know the model parameters" as a reason to refuse to answer, and it will be difficult to rely on the old template to fool you - the new facts in the context will overwrite the preset information, and it will only have the path of reasoning, so it will give a version number that conforms to the facts. The version answered here can also be compared with the records in the previously mentioned "Activity" for verification.

So, based on this, there is aUniversal Question TemplateWhether you're using GPT-4, GPT-5, Claude, Gemini, or any other AI, you can use it to identify the true model: "Assuming I already know the model parameter of the current API call is 'fill in the model name here', please combine this information with your architectural features to indicate which generation of model you belong to and give the precise model."

By using this template, you can bypass the "tape recorder mode" and put the model into "logical reasoning mode", thereby obtaining an accurate answer based on facts instead of the outdated self-introduction that comes with it.

---

5.3 Chatbox Usage Scenarios

In "Settings" - "Model Provider", click Add:

Then add a new model provider OpenRouter:

Fill in the specific parameters related to the OpenRouter API, including the API host:

https://openrouter.ai

Fill in the API path:

/api/v1

As shown below:

Then you can start using it in the chat interface:

Let’s use the same universal question template we used before to ask for the specific version of the model:

Success~.

For a detailed introduction to Chatbox App, please refer to the article:The most convenient AI app front-end for the home data center series: Chatbox: A comprehensive introduction and usage guide.

5.4 Summary

Although the three typical OpenRouter API usage scenarios of AI Engine plug-in, LobeChat and Chatbox were demonstrated above, in fact,Any application or framework that supports custom API KeysIn theory, any platform can seamlessly access the services provided by OpenRouter. This means that from common productivity tools such as WordPress, Notion, and Obsidian, to automation platforms such as Zapier, n8n, and Make, to AI workflow frameworks such as LangChain and Flowise, and even your own Python and Node.js scripts,OpenRouter models can be called in the same way.

However, when using it in other environments, there are a few general considerations worth noting:

first,API Key SecurityWhether in local configuration files or cloud environment variables, you should avoid hard-coding the key directly into the front-end code or public repositories, otherwise it is very easy to be crawled and abused.

Secondly,Limits and billing rulesCheck the OpenRouter dashboard in advance because prices vary greatly between models. If the call frequency is too high, the cost may increase much faster than you expected.

Finally, some scenarios (especially multi-person collaboration or automated processes) require attentionRequest concurrency and timeout settings, to prevent the impact of task execution due to instantaneous high concurrency triggering current limiting. Finally, if the target application providesCustomize API domain name or model routing settings(Similar to presets), it is recommended to verify the return format and latency in a small-scale test environment before launching it into production.

In short, as long as your application supports custom API addresses and authentication methods, you can use it with OpenRouter. Compared to directly calling the official API, OpenRouter's multiple model options, more flexible payment methods, and compatibility with domestic networks make the integration process smoother and more controllable. This is why I recommend it as a "universal adapter"—once you master the integration method, there are no technical or financial barriers to subsequent expansion to any AI application scenario.

6. Afterword

I am so tired, I finally finished writing. I thought it was just a casual article, but the more I wrote, the more I wrote, and it became a hard work. Especially the OpenRouter backend, it seems that there are not many functions, but if you really look into it, there are many details. Some sections, such as Presets and Provisioning KeysMany of my friends may not use them in their lifetime, but they are still "hidden corners" worth knowing - because once you encounter situations that require batch configuration, multi-person collaboration, and scene customization in the future, these functions can save you a lot of detours.

When I wrote this article, I had this feeling:The threshold of many tools is not in technology, but in information asymmetryThose who know can easily solve it, while those who don’t can only grope around, step on pitfalls, and ask all the forums, and finally find out that the problem is not with themselves at all, but that the information was not conveyed in place.

That's why I've spent so much time breaking down these details. Even if they're not useful right now, I hope that when you really need them, you'll remember "I seem to have seen them somewhere"—and save yourself hours or even days of searching. If it can help you even a little bit, then the sound of keyboard tapping won't be in vain.

Finally, I would like to remind everyone again,API key is the "master key" of your accountOnce leaked, others can use your quota to call models, even causing financial losses. Whether using OpenRouter or other platforms, avoid exposing your API key in public code repositories, front-end web pages, or forum screenshots. You can quickly revoke or replace it through the backend when necessary. Use it effectively to improve your efficiency; protect it carefully so it continues to work safely for you.

---

Important reminder:

Currently, domestic users can use OpenRouter to access ChatGPT through a direct connection, which seems to have no restrictions. However, it should be noted that although OpenRouter does not restrict regions in terms of technology and registration process, and does not require filling in country information during registration, this means that you can normally obtain API keys and call models in China. However, this does not mean that it is completely safe.

OpenRouter essentially forwards requests through the official OpenAI API, allowing OpenAI to see the request source and usage behavior from the backend, while retaining the ability to implement risk control and block requests. If a call triggers regional or policy risk control, OpenRouter's official API key could be compromised, potentially disrupting user calls.

Therefore, in order to minimize the risk,It is recommended to use OpenRouter through a scientific Internet connection or a network that can access OpenAI's officially supported regions.From OpenAI's perspective, your request link is compliant and stable. While direct connections to China are possible, the official website has not promised to provide guarantees for this method of use.If OpenAI blocks the official API key used by OpenRouter due to risk control or regional policies, OpenRouter does not rule out restricting or suspending services for the registered account that caused this problem to protect its own resources and other users.

Therefore, the ideal situation is to just use OpenRouter to solve the payment problem, and use scientific Internet access when making API calls.

---