I started using DeepSeek’s API about three weeks ago. Not because I needed another AI tool — I already pay for a few — but because a friend kept mentioning how cheap it was. I figured I’d burn a few bucks, test it out, and forget about it.
I didn’t forget about it.
But I also ran into things that weren’t obvious from the documentation. Some of them were annoying. Some were just… weird. And a couple genuinely surprised me. Here’s what actually happened when I tried to use the DeepSeek API for a small project.
The registration thing that almost stopped me
I signed up with my work email — the one I use for everything. Registration failed. The error said something like “this email domain is not supported.”
Wait, what?
I tried again. Same thing. I used my personal Gmail instead, and it worked immediately. No explanation on the site, no suggestion to switch. Just a dead end.
Turns out, DeepSeek blocks certain email domains. They don’t publish a list, but I’ve seen people mention this in forums. If you run into this, just switch to a Gmail or Outlook address. Don’t waste time trying to figure out why.
Authentication: personal vs. business
I went through the verification process without thinking too much. Personal ID, photo, done. But I noticed there’s also a business verification option. Out of curiosity, I looked into what changes if you go that route.
Honestly? Not much.
The features and pricing are identical. The only real difference is the paperwork — business verification requires your company registration info. And here’s the kicker: once you verify as a business account, you can’t switch back to personal. Ever. So if you’re just an individual developer testing things out, stick with personal verification. You’re not missing anything.
The money part
Recharging your balance is straightforward — Alipay or WeChat, depending on where you are. But if you’re outside China, that might not work. They also support wire transfers, but only for business accounts.
I put in about $15 to start. The balance doesn’t expire, which is nice. But here’s what I didn’t realize at first: they give you “bonus credits” sometimes, and those do expire. You can check the expiration date in the billing section. I almost lost a few dollars because I didn’t notice.
Invoicing takes about 7 business days if you need one. And the invoice name has to match your verification name. So if you verified as a person, you can’t get a company invoice. Makes sense, but worth knowing upfront.
The API behavior that confused me
I set up a simple test script — send a prompt, get a response. Standard stuff. I hit the endpoint and… nothing came back for about 20 seconds. Then suddenly the whole response appeared at once.
I thought I broke something.
Turns out, the web version uses streaming mode by default — you see each word as it’s generated. The API uses non-streaming by default. That means you wait for the entire response, then get it all at once. You can turn on streaming with stream=true. I did, and it felt way more responsive.
Also, I noticed some empty lines in the response during long waits. Those are keep-alive signals — the server sends blank lines so the connection doesn’t time out. If you’re parsing the response yourself, make sure your code ignores them, or they’ll mess up your output.
Rate limits and errors
I hit a 503 error on my second day of testing. Not because I was sending too many requests — I was barely sending any. It was just a bad time on their end.
DeepSeek doesn’t have hard concurrent limits per user. But when the system is busy, they do dynamic throttling based on your recent usage and overall load. So you might get 429 or 503 errors even if you’re not abusing it.
I was annoyed at first. But honestly? Every API has this problem during peak hours. I just retried after a few seconds and it worked.
One thing I wish I knew earlier: they don’t boost your concurrent limit if you ask. It’s the same for everyone. So if you’re planning to scale something big, you’ll need to design around that.
Token counting — it’s not obvious
I assumed token counting would be straightforward. 1 token ≈ 1 word, right?
Not really.
For English text, about 1 character equals 0.3 tokens. For Chinese, 1 character equals about 0.6 tokens. But this varies by model. The only reliable way to check is to look at the usage field in the API response.
They do provide a tool to calculate tokens offline. I downloaded the tokenizer package and tested it. It’s accurate enough, but honestly, I stopped worrying about it after the first few calls. The cost was so low that I wasn’t counting every token.
Oh, and they support LangChain. There’s a sample script in their documentation that replaces the OpenAI API key with the DeepSeek one. It took me five minutes to get it working.
What I actually think
Here’s the short version: DeepSeek’s API is cheap and works well. But it’s not perfect.
The email domain restriction is annoying. The non-streaming default caught me off guard. And the dynamic throttling means you can’t always predict when a request will fail.
That said, I’m still using it. For the price, it’s hard to beat. I’ve been running a small automation that calls the API about 200 times a day. It costs me around $0.50 per day. That’s absurdly low compared to other providers.
Would I recommend it? If you’re building something that doesn’t need ultra-low latency or guaranteed uptime, absolutely. If you’re building a mission-critical app that needs rock-solid reliability? Maybe not yet.
I’ll keep testing it and see how it holds up over a few months. Might change my mind. Might not.
Meta Description
DeepSeek API review: email domain issues, billing quirks, streaming vs non-streaming, rate limits, and token counting. Here’s what actually happened.
