Codex, Claude, and DeepSeek: How I Actually Use AI Tools from China
This is not a lab benchmark. It is my real tool split as a non-programmer in China building an export website: Codex for daily execution, Claude connected to DeepSeek for low-cost document work, and DeepSeek for fast rough work.
Many people online say Claude has stronger overall ability and more complete thinking.
Maybe they are right.
But I cannot honestly verify that claim in the same way they can, because I am in China and I cannot easily open and use a normal Claude paid membership. That changes the comparison. My question is not:
Which AI model is the strongest in theory?
My question is more practical:
With my access, my cost, my documents, and my export website work, which tool should do which job?
That is the comparison I care about.
My Current Short Answer
Right now, my most used setup is:
- Codex native subscription for real project execution, file editing, validation, and website work.
- Claude or Claude Code connected to DeepSeek for document processing, structure, and low-cost heavy reading.
- DeepSeek directly for fast rough drafts, Chinese thinking, and cheap first-pass work.
- DeepSeek connected to a vision model API when I need image understanding from Doubao, Qwen, or another visual model.
- Reasonix or command line when I want to use DeepSeek more directly instead of forcing it into Codex.
This is not because one tool is always better. It is because each tool has a different place in my workflow.
Claude With DeepSeek: Surprisingly Good for Documents
I have not been able to test a full Claude paid membership in the normal way. So I do not want to repeat other people's claims as if they are my own experience.
But I did use Claude or Claude Code connected to DeepSeek, and that surprised me.
When processing images, screenshots, or large document-like material, I found that it did not always need an expensive full visual model to be useful. With the right workflow, it could still analyze structure, extract meaning, and handle document tasks very well.
The strongest example for me was a large PDF.
I processed a PDF document of about 500 pages, and the total cost was less than 1 RMB in my use case. That was important. A task that might feel expensive with original frontier models became cheap enough to try without fear.
This is where Claude connected to DeepSeek became useful for me:
- large PDF reading
- document structure analysis
- image-like document understanding
- extracting useful sections from long material
- turning messy reference material into something I can act on
For this kind of work, cost changes behavior. If every test is expensive, I hesitate. If a 500-page document can be processed for less than 1 RMB, I can experiment more freely.
Codex: My Main Working Tool
Codex is the AI tool I use most often.
At first, I did not use it with a native subscription. I connected third-party APIs. In that setup, I felt Codex could be cheaper than some Claude-side workflows, especially when using third-party access to original models or API routes.
Later, I used Codex more as a native working environment. That changed my view. Codex is not only a chat model. It is useful because it can work inside the actual project.
For my website work, Codex is strong in three areas.
1. It Thinks Through the Work
When I give Codex a rough idea, it can usually turn that idea into a practical plan: what page to edit, what file to check, what risk exists, and what validation should run before calling the work finished.
That matters because I am not a programmer. I need the tool to help me translate business intent into executable work.
2. It Works With Real Files
This is the biggest difference from ordinary chat.
Codex can inspect the local project, edit HTML, update links, check sitemap and RSS files, run local servers, use Playwright, and create upload packages. It turns a suggestion into files I can actually upload.
For jjradiator.com and this site, that is more important than a beautiful answer.
3. Skills and Plugins Matter
To use Codex well, I think you have to use its skills and plugins well.
When Codex has the right skill for a task, it becomes more than a general assistant. It can follow a workflow: document polishing, website building, PDF handling, browser checking, customer outreach, or other specific work.
That is why Codex became my daily main tool. It is not only about model intelligence. It is about the working environment around the model.
DeepSeek: Very Cheap, Very Fast, but Rough
DeepSeek is the tool I respect most for cost.
It is extremely cheap. In my experience, if I compare rough usage cost, DeepSeek can feel like around one twentieth of a Claude Code original-model workflow and around one tenth of a Codex native workflow. This is not a formal benchmark. It is my practical cost feeling from real use.
DeepSeek is also very fast for me in China. Many replies arrive in milliseconds or a few seconds. That speed changes the way I use it. I can ask more questions, create more rough drafts, and explore ideas without worrying too much about cost.
But there is a problem.
The output often feels rough.
It is good at doing the first layer of work: summarize, draft, classify, expand an idea, produce a basic structure, or handle a low-cost first pass. But the writing can feel too simple, too mechanical, or too much like base-level labor.
That does not make it useless. It means I should not use it as the final voice.
My usual pattern is:
- Use DeepSeek to create a cheap first version.
- Use Codex, Claude-style workflows, or another stronger model to refine it.
- Use my own judgment to decide what is true, useful, and publishable.
DeepSeek is excellent for starting. It is not always enough for finishing.
How I Handle Images With DeepSeek
One important detail: I do not treat DeepSeek as the only model for every task.
When I need image recognition, I can let the DeepSeek workflow call another model API that is stronger at vision. For example, I can connect it to Doubao through Volcano Engine, or to a Qwen visual model, and use that model's image understanding ability.
That combination is very practical:
- DeepSeek stays as the cheap and fast thinking layer.
- Doubao, Qwen, or another visual model handles the image recognition.
- The final result comes back into the workflow for analysis, summarizing, or drafting.
This is useful because I do not need one model to do everything. I can let each model do the part where it is strongest.
For image-heavy documents, screenshots, catalog pages, or product references, this mixed workflow can be much better than asking a text-focused model to pretend it can see everything by itself.
Why I Do Not Like DeepSeek Inside Codex
I also tried connecting DeepSeek inside Codex.
My feeling was not good.
It did not feel as useful as Codex with its native model stack. Codex depends heavily on planning, tool use, file editing, and verification. When I use a weaker or rougher model inside that environment, the whole workflow feels less reliable.
So my current advice to myself is simple:
Do not force DeepSeek into Codex when Codex native models do the work better.
If I want to use DeepSeek, I would rather use Reasonix, PowerShell, CMD, or a direct command-line workflow. That keeps DeepSeek in the role where it is strongest: cheap, fast, direct processing.
My Practical Tool Split
| Tool | Where I Use It | Main Risk |
|---|---|---|
| Codex native | Website files, code edits, validation, upload packages, project execution | Higher cost than DeepSeek, so use it for work that needs reliability |
| Claude / Claude Code with DeepSeek | Large documents, PDF analysis, structure, low-cost heavy reading | Not the same as testing full Claude paid membership |
| DeepSeek direct | Cheap rough drafts, Chinese thinking, summaries, first-pass analysis | Output can be rough and needs polishing |
| DeepSeek + Doubao / Qwen vision API | Image recognition, image-heavy documents, screenshots, catalog pages | Needs API setup and careful checking of visual extraction |
| Reasonix / PowerShell / CMD | Direct DeepSeek agent or command-line workflows | Requires more comfort with terminal-style work |
This table is my real workflow, not a universal ranking.
My Current Recommendation
If I had to give one recommendation based on my own work, it would be this:
Use Codex for execution, DeepSeek for cheap rough work, Claude-style workflows for structure and document-heavy thinking, and a separate vision API when images matter.
For my daily work, Codex is still the main tool because it can actually touch the project. It can edit files, run checks, inspect pages, and close the loop.
DeepSeek is the best tool when cost and speed matter more than polish. When images matter, I prefer to connect it with a real vision model such as Doubao or Qwen instead of pretending DeepSeek alone should do everything.
Claude or Claude Code connected to DeepSeek is useful when I need a low-cost way to process large documents or organize complex material.
And if I need to use DeepSeek seriously, I prefer not to force it into Codex. I would use Reasonix or direct terminal workflows instead.
The Main Lesson
The best AI tool is not only the strongest model.
For a small operator, the best tool is the one that fits the job, the budget, the local access conditions, and the verification workflow.
That is why I do not want to write a simple "Claude vs Codex vs DeepSeek" ranking. My real conclusion is more practical:
- Codex is my main working environment.
- DeepSeek is my cheapest and fastest rough worker.
- For image tasks, DeepSeek works better when paired with Doubao, Qwen, or another vision API.
- Claude-style workflows with DeepSeek are useful for heavy document processing.
- The final result still needs human judgment and verification.
That is the tool split that currently helps me build this site and the export website behind it.
What Is Your AI Tool Split?
I want to compare this with other real workflows, especially from people who also care about cost, access limits, document processing, and practical execution.
If you use a different setup, send me your current split. I may summarize useful patterns in a future article, without publishing private details.
→ Previous article — why AI can make you less confident in a useful way.
→ Next article — how I used AI and TypeScript to cut a large PDF catalog into product images.