A Single Monthly Credit Pool Now Feeds 11 Chinese LLMs Into Claude Code and Codex

After using AI programming tools, the first problem I encountered wasn't insufficient models, but increasingly fragmented subscriptions.

Want to use GLM? Need to open one plan. Halfway through a project, feel Kimi is better for long context? Need to subscribe to another service. Switch to DeepSeek or MiniMax, and quotas, validity periods, and API Keys all need to be managed separately. What bothers me even more is the short-cycle quota window: even though the monthly fee is already paid, after coding continuously for a while, I have to stop and wait for the 5-hour window to recover, and then continue to keep an eye on the weekly quota. Taking Kimi's current official rules as an example, Kimi Code, besides sharing the monthly quota pool, also has separate 5-hour and weekly usage limits.

The AiiOnly Token Plan I've been using recently takes a different approach: instead of subscribing to a single model, it puts 11 mainstream domestic models into the same monthly Credits pool. The Standard domestic version's launch price is 79 yuan/month, including 6,320,000 Credits, which is 2.9 times the usage of the Lite plan.

11 Models in a Single Monthly Pool

Standard currently includes four series: DeepSeek, GLM, Kimi, and MiniMax, totaling 11 models. What attracts me most here isn't the quantity itself, but not having to bet in advance on 'which one I can only use this month.' Complex backends can be handed to models leaning towards reasoning and engineering capabilities, frontend pages can switch to a model better at long context or visual understanding, and testing, documentation, and minor modifications can use a faster-responding version.

The plan uses a single API Key and a unified Credits pool. The console provides both an OpenAI-compatible interface and an Anthropic-compatible interface:

OpenAI：https://llm.aiionly.com/v1/chat/completions
Anthropic：https://llm.aiionly.com/v1/messages

'Providing both types of interfaces' doesn't mean every model can use both protocols indiscriminately. The Anthropic /v1/messages endpoint only applies to some models on the platform that have been adapted to the Messages format. In actual integration, you still need to choose the protocol based on the model details and tool tutorials. OpenAI-compatible tools usually use /v1/chat/completions, while tools like Claude Code are more suitable for the adapted /v1/messages.

The console's model details provide the exact API model IDs; Standard's QPS is 2. When copying the model ID, it's best to take the value directly from here to avoid using the page display name as a request parameter.

The first time I checked the console, it still had the 10 models above. Recently, when I opened the model details again, GLM-5.3 had been added to the list, with the API model ID glm-5.3, so Standard currently has 11 models. It's not a replacement for GLM-5.2; both versions are still retained in the plan.

GLM-5.3 continues the foundation of GLM-5.2, mainly strengthening complex Coding and long-range tasks. Zhipu stated in its release notes that its internal Code Bench improved by 50% compared to GLM-5.2; this is the vendor's self-test result, and I haven't run a separate comparison on the same project yet. At this stage, I'll try it in large projects, complex debugging, and tasks requiring continuous Agent execution; stable workflows can continue using GLM-5.2.

These 11 models are not simple like-for-like replacements. Based on each vendor's official positioning, after placing them into an actual development workflow, my selection roughly looks like this:

Model	More Suitable Tasks	How I Would Use It
GLM-5.3	Strengthened complex Coding and long-range Agent tasks based on GLM-5.2	Large projects, complex debugging, and engineering tasks requiring continuous planning and execution
DeepSeek-V4-Pro	1M context, complex reasoning, Coding, and Agent long tasks	Backend architecture, core business logic, difficult bugs, and cross-module modifications
DeepSeek-V4-Flash	Also supports 1M context, lighter model, more flexible response for daily tasks	Quick code changes, code reviews, supplementing tests, and medium-complexity tasks
GLM-5.2	1M context, oriented towards project-level engineering and long-range Coding Agents	Large repository understanding, cross-file refactoring, complete delivery from requirements to deployment
GLM-5.1	200K context, emphasizing long-range execution, tool calling, and engineering optimization	Multi-stage development, performance optimization, Agent tasks requiring continuous iteration
Kimi-K3	1M context, native multimodal, long-range Coding, and knowledge work	Large projects, full-stack tasks, frontend implementation, and long document processing
Kimi-K2.7-Code	Dedicated programming model, stronger long-context instruction following and code task success rate	Daily coding, multi-file modifications, terminal tool collaboration
Kimi-K2.6	Balanced general capabilities, covering code, Agent, text, and visual input	General development, code Q&A, mixed Agent and visual understanding tasks
Kimi-K2.5	Native multimodal, supports thinking and non-thinking modes	Understanding pages from images, design-to-code, visual Agents
MiniMax-M3	1M context, native multimodal, Coding, and complex Agents	Long-cycle engineering tasks, large codebases, work requiring multi-step tool calls
MiniMax-M2.5	Coding, search, tool calling, and office tasks, higher execution efficiency	High-frequency coding, search organization, testing, documentation, and routine automation assistance

The table above describes the official capabilities of the models themselves. The fact that an original model supports images or videos does not mean that every current AiiOnly Token Plan interface has opened the corresponding input format; multimodal tasks still need to be based on the platform's interface documentation and actual testing.

Quota Managed Monthly, No Longer Interrupted by Short Windows

The Token Plan console displays the total monthly Credits and plan validity period, without breaking it down into 5-hour rolling windows or weekly quotas. For me, this is much more intuitive than 'recovering a portion every few hours': how much quota is left this month, and which model is consuming it quickly, can all be seen clearly on one page.

Credits are not just another way of writing Tokens, but the platform's unified quota unit. After a request is completed, the system deducts based on the actual input, output, and cache Tokens consumed by the specific model, so different models and tasks of different lengths cannot be compared solely by 'number of requests'.

This is also why I don't recommend describing it as 'unlimited calls.' Standard has 6.32 million Credits and QPS 2, and model multipliers also differ. It solves the frequent interruptions caused by short-cycle windows, not the elimination of all quota and concurrency limits. For a single person developing in interactive programming tools like Claude Code, Codex, or OpenCode, a monthly total pool is easier to arrange; high-concurrency batch tasks are not the use case for this plan.

VeryClaw Integrates Model Access and Chat Together

Besides connecting the Token Plan to existing programming tools, AiiOnly is also developing its own desktop client, VeryClaw. My first impression is that it feels a bit like WorkBuddy, but the entry points are more centralized: on the left, besides new conversations, there are Models, Agents, Channels, Skills, Scheduled Tasks, and Smart Assistants.

Model access is completed within the client. When adding an AI provider, select AiiOnly, enter the plan's API Key and model ID, and the models from the Token Plan can be added to VeryClaw. The page also notes that the API key is saved locally, eliminating the need to separately configure a relay tool just to start a conversation.

After connecting the model, you can directly return to the Chat conversation and select the model you just added below the input box. I used GLM-5.2 to propose a simple Python programming task, asking it to write a slider algorithm; the conversation not only returned the implementation idea and how to run it, but also showed the tool calls and the newly created slider_solver.py file. Model quota, conversation, tool execution, and file results all stay within the same client, which is closer to an actual workflow than just providing a chat box.

VeryClaw also places skills into the Chat input area; currently, entries for Word, PDF, PPT, skill lookup, and self-improvement are visible. Coupled with Agents, Channels, Scheduled Tasks, and Smart Assistants, its one-stop path is already clear: first access the Token Plan's models, then assign tasks in Chat, call skills or Agents when needed, and finally output results into documents and code files. The local gateway status at the bottom is used to confirm whether the execution environment is connected.

From adding models, starting conversations, to calling tools and generating files, VeryClaw has already placed several previously scattered steps into the same workbench. The Token Plan provides unified model quotas, and VeryClaw handles conversations, Agents, skills, and task execution; the entire usage process doesn't require switching back and forth between multiple clients.

First Connect Claude Code, Assign Different Models to Four Roles

AiiOnly's documentation center has already organized tutorials by tool. Besides Claude Code, you can also see entries for CC Switch, Cherry Studio, OpenCode, OpenClaw, Trae, Chatbox, WorkBuddy, and others. For the first connection, you don't need to guess environment variables and interface paths yourself; just follow the corresponding tool's page first.

I first added AiiOnly in CC Switch. For Claude Code's request address, fill in the service endpoint https://llm.aiionly.com; do not append the full /v1/messages path again. The API format selects Anthropic Messages, and the authentication field uses ANTHROPIC_AUTH_TOKEN.

Claude Code's model menu has role slots like Sonnet, Opus, Fable, and Haiku. I mapped them to DeepSeek-V4-Pro, DeepSeek-V4-Flash, GLM-5.2, and Kimi-K3 respectively. This way, switching roles in Claude Code actually sends the request to the actual model ID configured on the right.

There are two configuration details here that are easy to confuse. CC Switch's 'Request Address' requires the service endpoint, so fill in https://llm.aiionly.com; the https://llm.aiionly.com/v1/messages provided by the platform console is the full interface used for actual requests. If the tool automatically concatenates paths, manually entering /v1/messages again could cause address duplication. The model mapping on the right should also use the real ID given by the console, such as deepseek-v4-pro, and not just the Chinese or uppercase display name from the plan page.

Before saving, I first ran a connectivity test, and CC Switch returned aiionly connected successfully (29ms). This number reflects the interface connection latency at that time, not that the model generated a complete answer in only 29ms, but it at least quickly ruled out basic issues like unreachable addresses, network interruptions, and completely invalid authentication information.

Even after a successful connection, you still need to run a real conversation, because the speed test doesn't verify that every model supports the current protocol. The address might be accessible, but if the model isn't adapted for Anthropic Messages, the model ID is wrong, or the tool's attached request fields are incompatible, errors can still occur during the generation phase. My checking order is to first confirm the service endpoint, then verify the API format and model ID, and finally test the returned content with the simplest greeting. This way, when encountering problems, you can determine whether it's at the connection layer, protocol layer, or model layer, without repeatedly changing Keys and hoping for the best.

After re-entering Claude Code, the model menu already showed the four actual names configured earlier. The benefit here is direct: within the same project, without exiting the tool or switching accounts, you can switch between DeepSeek, GLM, and Kimi using /model. The four slots are just mappings in the current Claude Code menu and don't mean the plan can only use four models; the remaining models can be reconfigured to corresponding roles as needed by the task.

The [1m] in the menu is the context capability marker declared by CC Switch to Claude Code. Whether it's truly suitable to dump an entire ultra-long project in at once still needs to be judged based on the model's own context specifications, platform limits, and Credits consumption. For the first verification, there's no need to directly feed in the entire repository. I chose DeepSeek-V4-Flash for a minimal conversation test, first confirming that the request could reach the correct model from Claude Code and return parsable content.

After the model replied normally, the top of Claude Code also showed that deepseek-v4-flash[1m] was currently in use. A simple 'Hello' certainly doesn't prove it can handle complex engineering, but it confirms that the Anthropic Messages interface, CC Switch mapping, and plan Key are connected. Once this basic path is stable, letting the model read the project and execute tools will have much lower troubleshooting costs.

During backend development, I prioritize handing core architecture and complex logic to DeepSeek-V4-Pro, switching to Flash for daily modifications; when needing to read more project files at once, I switch to GLM-5.2 or Kimi-K3. It doesn't require each project to be bound to a single model; the division of labor can change with the task.

Then Connect Codex, Let Kimi-K3 Handle Frontend Tasks

Next, I connected the same plan to Codex. Here, the OpenAI Chat Completions protocol is used, not the Responses API. Codex's menu still shows its familiar model slots; in the configuration, I mapped these slots to AiiOnly's actual models: DeepSeek-V4-Pro, Kimi-K3, and GLM-5.2.

This time, I chose the gpt-5.4 slot, which actually corresponds to Kimi-K3. I asked it to create a native HTML, CSS, JavaScript 'AI Programming Task Board' in an empty directory, including three tasks, status filtering, progress switching, and responsive layout, without using build tools or frontend frameworks.

After receiving the task, Codex directly created index.html, style.css, and script.js. From the task request to starting to write files, everything stayed within the same session, and the current model slot was also displayed above the task.

The generated page can be opened directly in a browser; the filter buttons, task statuses, and overall progress are all interactive. The page cards showing GPT-5, Claude Sonnet 4, and Gemini 2.5 Pro are static task data for this demo board, not the actual models called by this request; the actual model that completed the page generation was the previously mapped Kimi-K3.

Returning to the Token Plan's model details, Kimi-K3 had 1 request, text input of about 0.013921M Tokens, text output of about 0.01322M Tokens, consuming 25,884.02 Credits. This set of data connects the tool side and the platform side: a frontend task completed in Codex indeed entered the same Standard monthly pool.

At this point, both sets of tools are fully operational. Claude Code uses DeepSeek, Codex uses Kimi-K3, and consumption is deducted from the same plan. In the future, if I want to switch to the newly added GLM-5.3, I just need to change the model mapping.

Final Reminders

Standard has 6,320,000 Credits per month, QPS is 2, and different models have different deduction multipliers; it's not unlimited use. The plan does not support refunds or downgrades after purchase. The Token Plan is only for use with interactive agents and programming tools like OpenClaw and Claude Code; you cannot use the plan Key to run custom application backends, automation scripts, or batch calls.

For me, managing just this one quota per month is much more comfortable than keeping an eye on several memberships simultaneously.