Take a Notification domain — the template/rules/preferences side and the thing actually pushing millions of messages out the door have completely different needs. But I never know when to actually split them. Do you wait until the data plane's scaling starts making deploys painful? Or is there some earlier "oh this is getting weird" signal I should watch for?
I’ve been using Qwen3-72B (and even the smaller 32B model) extensively alongside Opus 4.6 for the past few weeks.
My takeaway is that while Opus 4.6 still holds the edge in high-level system design and 'knowing what it doesn't know,' Qwen3 is remarkably competitive in Python and Rust implementations. In fact, for specific LeetCode-style logic or boilerplate generation, Qwen3 often feels faster and more concise.
The real strength of Qwen3 is its instruction-following capability—it rarely hallucinates API signatures compared to other open-source models I've tried. If you have the hardware to run the 72B (or even a high-quantized version), it's probably the closest you'll get to the Opus experience in the open-source world right now.
Fair point. And honestly, with more non-technical builders shipping agent-based products these days, that's probably where a service like this makes the most sense – for people who don't yet have the experience to know what guardrails to put in place.
Honestly, if you're designing your agent workflows properly with hard limits on retries and tool calls, the variance shouldn't be that wild. Most of the unpredictability comes from not having those guardrails in place early on. A few weeks of real production data usually shows the average cost is more stable than you'd expect.