I Was Debugging Timeout Errors at 2 AM, So I Tried Routing to Qwen3.8-Max

TL;DR: My side project's LLM calls kept timing out during a traffic spike because I only had one provider configured. I added a second model as fallback through RouteAI, and Qwen3.8-Max was one of the options — this is what that migration actually looked like, warts included.

The problem

I run a small tool that summarizes GitHub PRs for a Discord bot — nothing fancy, maybe 200-300 requests a day. Fine, until one evening a repo I follow had a burst of PR activity and my single LLM provider started returning 429s. My bot just... stopped responding. No fallback, no retry logic worth mentioning. Classic "worked fine until it didn't."

What I actually tried first

My first instinct was to just add retry-with-backoff. That helped a little, but if the provider itself is degraded, retrying into the same wall doesn't do much. I needed a second model I could fail over to, and I didn't want to write a second SDK integration from scratch for "just in case" code that mostly sits idle.

Where RouteAI came in

I'd seen RouteAI mentioned in a thread here on DEV a while back, so I gave it a shot mainly because it let me add a fallback model without rewriting my request logic. I configured Qwen3.8-Max as the secondary model. Setup itself took maybe 20 minutes, most of which was me re-reading my own retry code to understand where to slot the fallback call.

try:
    response = client.chat(model="primary-model", messages=messages)
except RateLimitError:
    response = client.chat(model="qwen3.8-max", messages=messages)

Yes, it's basically that simple in my case — your mileage will vary depending on how tightly coupled your prompts are to a specific model's quirks.

What I noticed (and didn't verify rigorously)
Response formatting was slightly different enough that I had to adjust my parsing regex once. Not a big deal, but worth knowing before you assume drop-in compatibility.
I didn't do a formal cost comparison — my volume is too low for that to be meaningful. If you're at higher volume, don't take my word for pricing, check current numbers yourself.
I haven't stress-tested this setup under a real outage yet, just the original traffic spike scenario. So "does the fallback actually save me next time" is still an open question for me.
Takeaway

This isn't a "this changed my life" post — it's a small annoyance I fixed on a side project. If you're running anything LLM-backed with real users depending on it, having a fallback model configured (regardless of which router or provider you use) is probably worth the 20 minutes it took me.