With one API for every model, you can route intelligently instead of picking one model for everything.
Pattern 1: route by task type
The simplest router is a lookup in your code:
MODELS = {
"classify": "claude-haiku-5-5",
"chat": "claude-sonnet-4-6",
"code": "claude-sonnet-4-6",
"hard": "claude-opus-4-7",
}Each feature in your app knows what it is, so it picks its model.
Pattern 2: route by difficulty
Let a cheap model judge difficulty first:
def pick(question):
r = client.chat.completions.create(model="claude-haiku-5-5", temperature=0, max_tokens=5, messages=[
{"role": "user", "content": f"Is this request simple or complex? Reply with one word.\n\n{question}"}])
return "claude-opus-4-7" if "complex" in r.choices[0].message.content.lower() else "claude-sonnet-4-6"The routing call is tiny, so it costs almost nothing.
Pattern 3: fall back on failure
import openai
CHAIN = ["claude-sonnet-4-6", "gpt-5.6-luna", "gemini-3.5-flash"]
def robust_chat(messages):
last = None
for model in CHAIN:
try:
return client.chat.completions.create(model=model, messages=messages, timeout=60)
except (openai.APIConnectionError, openai.APITimeoutError, openai.RateLimitError, openai.InternalServerError) as e:
last = e
continue
raise lastOnly fall back on errors that are about availability (timeouts, 429, 5xx). A 400 means your request is wrong, and another model won't fix it.
Pattern 4: escalate on low confidence
Ask the first model to say ESCALATE when it isn't sure, then retry with a stronger model. This works well for support bots.
Measure
Log which model handled each request. Usage Logs show cost per model, so you can see whether the routing is actually saving money.
Claude, GPT, Gemini, DeepSeek, image and video models through one OpenAI- and Anthropic-compatible endpoint. Prepaid, pay only for what you use.