DEV Community

Daniel Dong
Daniel Dong

Posted on

I migrated my codebase from 1 AI provider to 4. Here's every line I changed.

#ai #api #llm #kimi

I migrated my codebase from 1 AI provider to 4. Here's every line I changed.

# Before
client = openai.OpenAI(
    api_key="sk-deepseek-xxx",
    base_url="https://api.deepseek.com/v1"
)
client.chat.completions.create(model="deepseek-chat", messages=msgs)

# After
client = openai.OpenAI(
    api_key="mb-xxx",
    base_url="https://aibridge-api.com/v1"
)
client.chat.completions.create(model="deepseek-chat", messages=msgs)
Enter fullscreen mode Exit fullscreen mode

That's the migration. One string in the base_url. One string in the
api_key. The import openai, the .create() signature, the response
parsing, the streaming logic, the error handling — all unchanged.

But now I can do this:

# Fast and cheap ($0.27/M) — 90% of traffic
client.chat.completions.create(model="deepseek-chat", messages=msgs)

# Multilingual, 128K context
client.chat.completions.create(model="qwen-max", messages=msgs)

# Complex Chinese reasoning
client.chat.completions.create(model="glm-4-plus", messages=msgs)

# 1M context, always reasoning
client.chat.completions.create(model="kimi-k3", messages=msgs)
Enter fullscreen mode Exit fullscreen mode

Four providers. Four different strengths. Same function call. Same
response format. No SDK switches. No new environment variables.

The routing pattern that saves money

Most apps don't need the most powerful model for every request. My
app does classification, summarization, simple generation, and the
occasional complex debugging session. Four types of work. Four models.

def pick_model(task):
    return (
        "deepseek-chat" if task in ("classify","summarize","simple")
        else "qwen-max" if task in ("translate",)
        else "glm-4-plus" if task in ("reason",)
        else "kimi-k3" if task in ("review_codebase",)
        else "deepseek-chat"
    )

client.chat.completions.create(model=pick_model(task), messages=msgs)
Enter fullscreen mode Exit fullscreen mode

Before routing: every request hit the 7/Mmodel.Monthlybill: 700.
After routing: 90% of requests hit the 0.27/Mmodel.Bill: 120.

Same quality on classification and summarization. The cheap model
is indistinguishable from the expensive one on those tasks. The only
difference is the price tag.

Kimi K3: the one worth upgrading for

When a task needs real reasoning power and a massive context window,
K3 is the only model I reach for. 1M token context — I paste entire
modules and ask it to trace race conditions through the call chain.
It reads every line, runs an automatic reasoning pass, then gives me
locations with suggested fixes.

No "think mode ON/OFF" toggle. No prompt engineering. No configuration.
Just send the prompt. K3 thinks, then answers.

The supporting cast

*Free playground *— 15 models, 10 requests/day, no signup. Try K3
without creating an account. If you like it, GitHub OAuth gets you a
key in five seconds.

Prompt library — 24 ready-to-use prompts across 6 categories.
Each one recommends the right model. "Find bugs → Kimi K3." "Translate
→ Qwen Max." One click fills the playground.

Onboarding — after login, your API key is on the screen with a
copy button. A curl command you can paste immediately. Python SDK
code. 30 seconds to first response.

Dashboard — usage bar that escalates. Polite below 50%. Orange at
80%. Red and flashing at 90%. The upgrade button gets harder to miss
exactly when you need to act.

Free tier: 500K tokens/month. No credit card. One import openai.
One endpoint. 15 models.

→ aibridge-api.com/playground.html
→ aibridge-api.com/prompts.html

1

2

3

4

Top comments (0)