Look, i Spent Weeks Comparing AI Models and the Prices Made Me Gasp
Okay so I just graduated from a coding bootcamp a few months ago, and like every other new dev out there, I started building projects with OpenAI's API. It felt normal. You sign up, you grab an API key, you send some requests, and you get charged whatever they charge. I didn't think twice about it.
Then one night I was doom-scrolling Reddit at 2am (we've all been there) and I stumbled into some thread about Chinese AI models. People were saying things like "DeepSeek is 40x cheaper than GPT-4o." I genuinely thought they were joking. Like, that's a huge claim. There's no way a model from China is matching OpenAI for a fraction of the price. Right?
Wrong. So wrong. What I found over the next few weeks of testing genuinely blew my mind, and I want to walk you through everything because I think a lot of bootcamp grads like me are getting absolutely played on pricing.
My First Big Surprise: The Cost Difference Is Real
Let me just throw some numbers at you because I stared at this spreadsheet for like an hour trying to make sense of it. These are the actual API prices per million tokens, straight from each provider's pricing page.
| Model | Where It's From | Input Price | Output Price |
|---|---|---|---|
| GPT-4o | United States | $2.50 | $10.00 |
| Claude 3.5 Sonnet | United States | $3.00 | $15.00 |
| Gemini 1.5 Pro | United States | $1.25 | $5.00 |
| GPT-4o-mini | United States | $0.15 | $0.60 |
| DeepSeek V4 Flash | China | $0.18 | $0.25 |
| Qwen3-32B | China | $0.18 | $0.28 |
| GLM-5 | China | $0.73 | $1.92 |
| Kimi K2.5 | China | $0.59 | $3.00 |
Read that DeepSeek row again. Output price is $0.25 per million tokens. GPT-4o is $10.00 per million tokens. That's not a typo. That is 40 times cheaper. I had no idea this was even possible.
For context on what this means in real life, when I built my first little chatbot project, I burned through about $14 in two weeks just playing around. With DeepSeek V4 Flash at those prices, I could've run the same chatbot for the entire year and probably spent less than a coffee.
But Wait, Are The Chinese Models Actually Good?
This was my next question. Cheap means nothing if the model is garbage. So I dug into benchmarks. I want to be upfront here, these numbers come from community testing and aren't perfectly scientific, but they tell a pretty consistent story.
Reasoning Tests (the MMLU-style stuff)
| Model | Score | Output Price |
|---|---|---|
| Claude 3.5 Sonnet | 89.0 | $15.00 |
| GPT-4o | 88.7 | $10.00 |
| Qwen3.5-397B | 87.5 | $2.34 |
| Kimi K2.5 | 87.0 | $3.00 |
| GLM-5 | 86.0 | $1.92 |
| DeepSeek V4 Flash | 85.5 | $0.25 |
Look at that DeepSeek row. A score of 85.5 versus GPT-4o's 88.7. So GPT-4o is maybe 3% smarter in raw reasoning benchmarks, but it costs 40x more. Let that sink in. You're paying 40x the price for a 3% improvement on a benchmark that may or may not even matter for what you're building.
Code Generation (HumanEval benchmark)
| Model | Score | Output Price |
|---|---|---|
| Claude 3.5 Sonnet | 93.0 | $15.00 |
| GPT-4o | 92.5 | $10.00 |
| DeepSeek V4 Flash | 92.0 | $0.25 |
| Qwen3-Coder-30B | 91.5 | $0.35 |
| DeepSeek Coder | 91.0 | $0.25 |
I was shocked when I saw this. DeepSeek V4 Flash actually scores 92.0 on HumanEval. That's basically tied with GPT-4o. And it's still that absurdly cheap $0.25 per million tokens price.
If you're a bootcamp grad building side projects like me, this should matter to you. A LOT.
Chinese Language Tasks (C-Eval benchmark)
| Model | Score | Output Price |
|---|---|---|
| GLM-5 | 91.0 | $1.92 |
| Kimi K2.5 | 90.5 | $3.00 |
| Qwen3-32B | 89.0 | $0.28 |
| GPT-4o | 88.5 | $10.00 |
| DeepSeek V4 Flash | 88.0 | $0.25 |
This part is wild. The Chinese models absolutely crush it on Chinese language tasks, which makes sense, they're literally built for it. But even GPT-4o lands at 88.5, which means for Chinese language work, you're paying $10.00 per million tokens for something that performs slightly worse than a model costing $0.25.
The Actual Problem Nobody Talks About
So at this point I was pretty excited. I figured I'd just sign up for DeepSeek or Qwen and start saving money immediately. That's when I hit the wall.
Try to sign up for DeepSeek. You need a Chinese phone number. Try to pay for Qwen. WeChat or Alipay only, which I don't have and you probably don't either. Try to read the docs for GLM. Mostly in Chinese. Try to access Kimi's API from outside China. Half the time it just doesn't work.
I spent an entire weekend trying to actually USE these cheap models and I couldn't get past step one. It was incredibly frustrating. Here's a quick comparison of the actual experience:
| What You Need | US Models | Chinese Models |
|---|---|---|
| Payment method | Credit card works fine | WeChat or Alipay only |
| Account creation | Email and done | Needs Chinese phone number |
| API format | OpenAI standard | Different for each provider |
| Works outside China | Yes everywhere | Often geo-blocked |
| Docs in English | Yes | Mostly Chinese |
| Support in English | Yes | Mostly Chinese |
| Charged in | Dollars | Yuan |
This is the part that bootcamp doesn't teach you. Knowing that a model is cheaper means nothing if you literally can't pay for it or sign up for it. I was about to give up on the whole Chinese AI thing when a friend mentioned Global API.
How I Actually Started Using These Models
Global API is basically a service that handles all the annoying international stuff for you. They give you an OpenAI-compatible endpoint, you pay with PayPal or a regular Visa card, and under the hood they're routing your requests to Chinese models like DeepSeek and Qwen. You don't need a Chinese phone number. You don't need to figure out Alipay. You just write code like normal.
I tried it and it worked on the first try, which almost never happens in my experience.
Here's a quick Python example showing how I set up my chatbot to use DeepSeek V4 Flash:
import openai
client = openai.OpenAI(
api_key="your-global-api-key",
base_url="https://global-apis.com/v1"
)
response = client.chat.completions.create(
model="deepseek-v4-flash",
messages=[
{"role": "system", "content": "You are a helpful coding assistant."},
{"role": "user", "content": "Write me a Python function to reverse a linked list."}
]
)
print(response.choices[0].message.content)
If you've used the OpenAI SDK before, this looks identical. That's the whole point. I literally copy-pasted my existing code, changed the base URL, swapped the model name, and it just worked. No new SDK to learn. No weird new authentication flow.
Here's another example where I switched between Qwen and DeepSeek for a simple task:
from openai import OpenAI
client = OpenAI(
api_key="your-global-api-key",
base_url="https://global-apis.com/v1"
)
# Try Qwen for general chat
qwen_response = client.chat.completions.create(
model="qwen3-32b",
messages=[{"role": "user", "content": "Explain async/await in simple terms"}]
)
# Try DeepSeek for code-heavy work
deepseek_response = client.chat.completions.create(
model="deepseek-v4-flash",
messages=[{"role": "user", "content": "Refactor this function to be more Pythonic"}]
)
print("Qwen says:", qwen_response.choices[0].message.content)
print("DeepSeek says:", deepseek_response.choices[0].message.content)
Switching models is just changing a string. I went from paying OpenAI rates to paying basically nothing in about 10 minutes. That's when I knew this was going to change how I build projects.
The Matchups That Actually Matter
Let me walk you through the most interesting comparisons I found, because if you're like me, you want to know exactly which model to swap in for which job.
DeepSeek V4 Flash vs GPT-4o
These two feel like the main event. DeepSeek is the cheapest of the bunch and GPT-4o is the most popular American model.
| What I Looked At | DeepSeek V4 Flash | GPT-4o | Who Wins |
|---|---|---|---|
| Output price | $0.25 per million | $10.00 per million | DeepSeek, by 40x |
| General quality | Very good | Slightly better | GPT-4o, barely |
| Code generation | Excellent | Excellent | Basically tied |
| Speed | 60 tokens per second | 50 tokens per second | DeepSeek |
| Context window | 128K tokens | 128K tokens | Tie |
| Can process images | No | Yes | GPT-4o |
For 95% of what I'm building, DeepSeek is the right call. The only place where GPT-4o genuinely beats it is image understanding. If your app needs to look at photos and describe them, you still need GPT-4o or another vision-capable model. But for chatbots, code helpers, content generators, summarizers, data extraction, basically all the text stuff, DeepSeek is the move.
Qwen3-32B vs GPT-4o-mini
This one made me laugh because Qwen wins on every single dimension I tested.
| What I Looked At | Qwen3-32B | GPT-4o-mini | Who Wins |
|---|---|---|---|
| Output price | $0.28 per million | $0.60 per million | Qwen, 2.1x cheaper |
| Overall quality | Good | Okay-ish | Qwen |
| Code generation | Good | Just okay | Qwen |
| Chinese language | Excellent | Mediocre | Qwen |
I genuinely cannot find a reason to use GPT-4o-mini over Qwen3-32B right now. The price is better, the quality is better, and the code generation is better. Bootcamp grads, please hear me when I say this: stop defaulting to GPT-4o-mini just because it's the easy OpenAI option.
Kimi K2.5 vs Claude 3.5 Sonnet
This is the matchup I was most curious about because Claude has a reputation for being really good at reasoning.
| What I Looked At | Kimi K2.5 | Claude 3.5 Sonnet | Who Wins |
|---|---|---|---|
| Output price | $3.00 per million | $15.00 per million | Kimi, 5x cheaper |
| Reasoning ability | Excellent | Excellent | Basically tied |
| Chinese language | Excellent | Mediocre | Kimi |
On pure reasoning tasks, they score basically the same. Claude's slight edge in some specific areas might matter for serious research projects, but for most of what I'm building, Kimi gets the job done for one-fifth the price.
My Honest Take After All This Testing
Here's the thing nobody tells you when you're learning to code. The expensive, popular American models are not 40 times better than the cheap Chinese ones. They're maybe 5% better on some benchmarks, and on others they're actually worse. The price gap is huge and the quality gap is tiny.
For a bootcamp grad like me who is building portfolio projects, side hustles, and trying not to bleed money on API costs while I'm still figuring things out, this changes everything. I can build a chatbot that costs me basically nothing to run. I can experiment with prompts without staring at my usage dashboard in horror.
That said, I want to be fair. There are still some places where the US models make sense:
- If you absolutely need vision capabilities, GPT-4o still wins
- If you're building something where 2-3% benchmark improvement actually matters to your business, the US models still have a small edge
- If you need the absolute best in conversational nuance and creative writing, Claude is still a solid choice
But for the vast majority of what most developers are building? The Chinese models are a no-brainer.
My Recommendation If You Want To Try This
If you've read this far and you're curious, here's what I'd suggest. Start by signing up for Global API. It took me about five minutes. You verify your email, add PayPal or a card, and you're in. They give you that OpenAI-compatible endpoint at global-apis.com/v1, so any code you've already written for OpenAI will work with minimal changes.
Try DeepSeek V4 Flash first. It's the cheapest one and the benchmarks are impressive. Build a small project with it. Watch how little you spend. Then try Qwen3-32B and Kimi K2.5 for different tasks. You'll get a feel for which model fits which job, and you'll save a ton of money in the process.
I'm not being paid to say this, I'm just a bootcamp grad who discovered something that actually helps me and wanted to share it. Check out Global API if you want to experiment with these models without dealing with all the international payment nonsense. It genuinely made
Top comments (0)