Last weekend, I hit 0% usage on my $100-a-month Codex plan.
I bought another $30 in credits. Those disappeared within two hours.
I still had projects I wanted to work on, so my choices seemed to be:
- Spend even more money.
- Wait for my limits to reset.
- Find another model.
Then I saw this tweet:
That was intriguing enough, so I had to try it.
Since I ran out of limits in Codex, I had to ask ChatGPT how to configure DeepSeek inside Codex (what a sentence). It gave me a long setup plan, and I followed it manually.
How I connected DeepSeek to Codex
Apparently, DeepSeek has official instructions on how to do it.
It recommends a one-click setup script, and I am lazy, but I didn’t want to run a mystery script against my development environment. So I did actually look at it.
It's just over 1000 lines and luckily these two lines were in the beginning:
In short, it looks like even after I downloaded the script, it'll still somehow install whatever version is at this CDN address at the type of script execution. ChatGPT confirmed my concerns and I decided to read Option 2.
It was short enough:
- Create a `~/.codex/models.json` file and copy/paste what they have on the website (I'll admit, I didn't read that one)
- Make a few updates to `~/.codex/config.toml`, including pasting the API key.
That was about it. Not really sure why a 1000+ lines script is needed at all.
I created an API key, added a couple of bucks to the account, and tested it.
Did it actually work?
Surprisingly, it worked on the first try. For my first test, I asked:
"What’s the next item in this project for me to do? It should have been recorded in one of the docs."
DeepSeek searched the project, found the directory containing my execution plans, and returned an answer based on them.
That doesn’t prove it can replace the models I normally use. It does prove the integration worked and that the model could inspect a real codebase through Codex instead of answering a toy prompt.
The real evaluation came in the next few days when I had it work on those actual tasks. It did work. Most of the time I wouldn't even know it wasn't the actual Codex model running. Until I received a notification about an alarm trigger.
I'll admit - I've been pretty terrible at reading the code AI produced. Especially for my personal projects. I also learned to somewhat trust Codex, especially since I have guardrails like making sure it deploys to sandbox and verifies that the implementation actually works, and then GitHub Actions run tests after the deployment to a beta branch in Amplify.
What happened was - DeepSeek really overcomplicated a DDB schema, didn't account for an edge case that would happen for every new guest user of the app, and didn't really test this scenario. The details aren't very important. What's important is to remember: cheap tokens are useful. Cheap tokens that produce bad work are still expensive.
Of course, it's mostly on me for not looking through the PRs in more detail. But it's also the only thing that I found that it did terribly wrong. It did succeed on smaller and less ambiguous tasks.
Why the price caught my attention
At the time of writing, DeepSeek V4 Flash costs $0.14 per million uncached input tokens and $0.28 per million output tokens. For comparison, OpenAI lists GPT-5.6 Sol at $5 per million input tokens and $30 per million output tokens.
That is not a claim that the models are equally capable. It also isn’t a direct comparison between API billing and my Codex subscription. It simply explains why the experiment was interesting.
In about a week, I spent less than $7 on DeepSeek. That includes me using it in OpenClaw as well. Compared to $30 I wasted on Codex's extra credits, it's not that bad.
If you want to watch the complete setup, I recorded the process here.
Cheers!
Evgeny Urubkov (@codevev)
