Every LLM app has the same hidden cost: it sends every request to the most expensive model, because that's the one it was built on.
It doesn't have to. Here's the fix, and the tooling that makes it one line instead of a migration.
The problem, quantified
A budget chat model costs ~ 0.27permillioninputtokens.Aflagshipreasoningmodelcosts 15 per million output tokens — and reasoning models burn most of their tokens in the thinking phase, which all bills as output.
If your app runs on the flagship model for everything, you're paying 50x for tasks that a budget model handles fine. A "summarize this email" doesn't need the same model as "debug this concurrency deadlock."
The fix: route by task, not by default
The single most valuable line in our stack is the one that picks a model per request:
def answer(q: str) -> str:
model = "deepseek-chat" if len(q) < 500 else "kimi-k3"
return client.chat.completions.create(
model=model,
messages=[{"role": "user", "content": q}],
).choices[0].message.content
Cheap model for the long tail of easy requests. Strong reasoning model for the hard 5%. The switch costs one string — which means you'll actually do it, instead of "planning to migrate someday."
Why one endpoint changes everything
Routing only works when the switching cost is zero. That's the whole point of AIBridge — an OpenAI-compatible gateway in front of 15+ models:
- One key, every model — deepseek-chat, deepseek-v4-flash, kimi-k3, glm-4-plus, qwen-plus, and more, all behind the same OpenAI SDK
- Streaming on all of them — no per-provider streaming quirks
- Embeddings through the same key — no second vendor to onboard
Then make the free tier do real work
Routing matters more when the budget is real. AIBridge gives you 500K tokens every month, reset automatically — enough to route all your easy requests through the cheap model and reserve the expensive one for what genuinely needs it.
When free runs out, pricing stays boring and predictable:
- Top-up packs: 1M tokens 2.99⋅5M9.90 · 20M $29.90 — raw 1:1, no expiry, no multiplier games
- Pro: $9.90/month for 5M monthly tokens + higher rate limits
The part nobody wants to build (so we did)
- Per-model usage on one dashboard — see which model eats your budget, then reroute it
- Failover & health tracking — if a provider degrades, traffic moves instead of piling up errors
- PII redaction — masked before requests leave your app
The one-line summary
Stop paying flagship prices for budget tasks. Route by task, switch models by changing a string, and start from a free tier that actually covers the cheap 95% of your traffic.
Top comments (0)