DEV Community

rarenode
rarenode

Posted on

I Cut My AI Bill 90% Testing These 4 Chinese Models

I gotta say, i Cut My AI Bill 90% Testing These 4 Chinese Models

I still remember the month my OpenAI bill hit $1,200. I was running a SaaS with three AI features, and every time a user touched one, I'd wince a little. That sting pushed me into a 30-day deep dive into Chinese models — DeepSeek, Qwen, Kimi, and GLM — all routed through Global API's unified endpoint so I could A/B test without juggling four dashboards.

Here's the thing: I wasn't trying to be a global citizen or make some statement about AI geopolitics. I just wanted my bill to stop hurting. What I found was honestly ridiculous. Some of these models cost literal pocket change compared to what I was paying, and a couple of them are flat-out better than GPT-4o for the things I needed.

Let me walk you through what I learned, the spreadsheets I burned through at 2 a.m., and where I landed.


The Setup: What I Was Spending Before

Before we get into the Chinese model showdown, I want to show you my "before" picture so the "after" makes sense.

My stack was mostly GPT-4o for chat features and a touch of GPT-4o-mini for the cheap stuff. At $10/M output tokens on GPT-4o and $0.60/M on mini, my burn rate was stupid. Check this out — I crunched the math on just one feature (an AI assistant that summarizes long documents):

  • Average output per request: ~400 tokens
  • Daily requests: ~3,000
  • Monthly tokens: 36M output tokens
  • Monthly cost on GPT-4o: $360
  • Same traffic on GPT-4o-mini: $21.60

That gap right there is why I went hunting. If a model gets me 90% of GPT-4o's quality for $0.25/M output instead of $10/M, I'm saving 97.5% on that single feature. That's wild.


DeepSeek V4 Flash: The New King of My Stack

Let me start with the model I ended up keeping on for most of my production traffic: DeepSeek V4 Flash at $0.25/M output tokens.

I'm calling it first because that's where the savings are. At $0.25/M, DeepSeek V4 Flash is 40x cheaper than GPT-4o. Forty times. I had to triple-check my math because it sounded like a typo.

The DeepSeek lineup has something for every budget tier:

Model Output $/M What I'd use it for
V4 Flash $0.25 Daily chat, content, light code
V3.2 $0.38 Newer architecture, similar role
V4 Pro $0.78 Production-grade when I need a quality bump
R1 (Reasoner) $2.50 Hard math, multi-step logic
Coder $0.25 Code-specific workloads

What hit me when I tested V4 Flash was the speed. I'm clocking around 60 tokens per second on average, which is among the fastest I've measured across any provider. The English quality holds up against Western standards — I'd put it neck-and-neck with GPT-4o on my document-summarization eval.

Where it loses points for me:

  • No real vision/multimodal story. If I need image understanding, DeepSeek isn't where I go.
  • Chinese-language nuance is slightly behind the specialists.
  • Fewer model size options than Qwen, so if I need fine-grained control over cost vs. quality, I'm a bit boxed in.

But honestly, for the price? It doesn't matter. Here's how I wired it up:

from openai import OpenAI

client = OpenAI(
    api_key="ga_xxxxxxxxxxxx",
    base_url="https://global-apis.com/v1"
)

response = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[
        {"role": "user", "content": "Summarize this support ticket thread in 3 bullets."}
    ],
    max_tokens=300
)

print(response.choices[0].message.content)
Enter fullscreen mode Exit fullscreen mode

That's literally my "expensive" route now. $0.25/M. I'm going to bed smiling.


Qwen: The One With Every Shape and Size

Here's the thing about Qwen — if Alibaba made a Swiss Army knife out of LLMs, this is what it'd look like. Their model range goes from $0.01/M all the way up to $3.20/M, which means there's literally a Qwen model for whatever weird thing I'm trying to do.

Let me show you what I mean:

Model Output $/M My use case
Qwen3-8B $0.01 Throwaway tasks, classification, spam filters
Qwen3-32B $0.28 My default general-purpose model
Qwen3-Coder-30B $0.35 When I need code that actually compiles
Qwen3-VL-32B $0.52 Reading images, PDFs, screenshots
Qwen3-Omni-30B $0.52 Mixed audio/video/image stuff
Qwen3.5-397B $2.34 The "I need adult supervision" tier
Qwen3.6-35B $1.00 (Honestly feels overpriced — I skip it)

That $0.01/M on Qwen3-8B is not a typo either. One cent per million output tokens. I started routing high-volume, low-stakes classification through it and my costs on that pipeline dropped to basically nothing.

The Qwen vision models (VL series) saved me when I needed to pull text out of uploaded screenshots — that feature alone is what made me keep an active Qwen integration alive. The Omni model handled a video summarization test I was running and did it cleanly.

Where Qwen annoys me:

  • The version names are chaotic. Qwen3 vs Qwen3.5 vs Qwen3.6, with 8B/30B/32B/35B/397B sprinkled in. I keep a sticky note on my monitor.
  • The mid-tier English isn't quite DeepSeek level — it's good, just not as crisp on long-form generation.
  • That $1/M Qwen3.6-35B line feels like a cash grab compared to the rest of the family.

For day-to-day general work, Qwen3-32B at $0.28/M became my second-favorite model. Here's the kind of thing I run through it:

response = client.chat.completions.create(
    model="Qwen/Qwen3-32B",
    messages=[
        {"role": "system", "content": "You write clear, idiomatic Python."},
        {"role": "user", "content": "Merge two sorted lists without using sort()."}
    ]
)
print(response.choices[0].message.content)
Enter fullscreen mode Exit fullscreen mode

That's $0.28/M for code that, in my testing, comes out cleaner than GPT-4o for routine stuff. I cannot overstate how strange it feels to write that sentence in 2026.


Kimi: The Brainy One I Can't Afford (Yet)

Kimi is the model I want to use more than I actually do, and that's strictly a price thing.

Kimi K2.5 sits at $3.00/M output. That's three dollars per million tokens. Compared to the sub-dollar options elsewhere, Kimi looks expensive on paper. But here's where I have to be honest: when I tested Kimi on the hardest reasoning prompts in my eval set — multi-hop logic, tricky math, anything that required holding three concepts in mind simultaneously — Kimi was noticeably sharper than the others.

The whole Kimi family is premium-priced:

Model Output $/M What I noticed
K2.5 $3.00 My reasoning benchmark leader
(Other Kimi models) up to $3.50 All premium, no budget option

That's the rub with Kimi. There's no "cheap Kimi." The top of the family starts at $3.00/M, which is 12x more expensive than DeepSeek V4 Flash. There's no small model to route traffic through while saving money.

What I ended up doing: I keep Kimi on standby for the 5% of requests where reasoning quality genuinely matters. Medical claim adjudication, complex legal summarization, that sort of work. For everything else, the cheaper models do the job and I keep Kimi's invoice low.

The scoring breakdown I'm seeing on benchmarks matches my gut:

  • Kimi: ⭐⭐⭐⭐⭐ on reasoning
  • Kimi: ⭐⭐⭐⭐⭐ on Chinese language
  • Kimi: ⭐⭐⭐ on speed (it's noticeably slower than V4 Flash)
  • Kimi: ❌ on vision/multimodal

If your workload is reasoning-heavy and you can stomach $3.00/M, Kimi is genuinely the best of these four at that specific job. If you're running chat features for a SaaS at scale, the math just doesn't work.


GLM: The Chinese-Language Specialist That Surprised Me

GLM came onto my radar late, and I almost skipped it because I assumed it was just "another Chinese model." I was wrong.

Zhipu AI's lineup has a surprisingly wide spread:

Model Output $/M My takeaway
GLM-4-9B $0.01 Tied for cheapest with Qwen3-8B
GLM-5 $1.92 The flagship, $1.92/M

That $0.01/M on GLM-4-9B is significant because GLM punches above its weight on Chinese-language tasks. If you're serving a Chinese user base — or your prompts come back in mixed English/Chinese — GLM often reads more natural than the others.

I tested GLM-5 on a benchmark of Chinese marketing copy rewrites and it edged out Kimi on natural phrasing. That's wild to me, given Kimi's reputation. The GLM-4.6V vision model handled some product-image description tests I ran, which gave me a viable second option to Qwen's VL lineup.

What I don't love:

  • GLM-5 at $1.92/M is the most expensive of the "flagship" tier from these four providers (Kimi at $3.00/M aside). I'm not always paying that premium for what I get back.
  • English performance is solid but not DeepSeek tier.

For me, GLM is now my "Chinese-heavy workflow" model. The combination of GLM-4-9B at $0.01/M for routine stuff and GLM-5 at $1.92/M for the nuanced Chinese-only jobs gives me a complete toolkit.


The Math That Actually Matters: A Side-by-Side Cost Test

I want to give you one concrete example that crystallized the decision for me. Same prompt, same output length, four different providers:

Prompt: "Write a 400-word product description for a noise-canceling headphone, optimized for SEO."

Output: ~400 tokens per request × 5,000 requests/month = 2M output tokens/month

Provider & Model Cost per 1M output Monthly cost
GPT-4o (my old setup) $10.00 $20.00
DeepSeek V4 Flash $0.25 $0.50
Qwen3-32B $0.28 $0.56
Kimi K2.5 $3.00 $6.00
GLM-5 $1.92 $3.84

That's a single feature. My monthly bill on that one product description pipeline went from $20.00 to $0.50. A 97.5% reduction. Across my whole app — chat, summarization, code assist, image captioning — I landed on a roughly 91% reduction in my AI spend.

Here's what my actual monthly stack looks like now:

-

Top comments (0)