I got the Qwen Pro plan to save some money on Claude credits. It was a mistake — and I want to be precise about what kind of mistake, because the two easy conclusions people jump to are both wrong.
The model is not the complaint
The easy wrong conclusion is that Qwen isn't good. It is good. Qwen 3.8 is great — in my own use, almost as good as Opus. That's not a hedge or a consolation prize I'm offering the model on its way out the door. That's my actual judgment after running real work through it. If someone asked me whether Qwen 3.8 can do serious agentic work, I'd say yes without qualification.
The other easy wrong conclusion is that I just don't understand how to manage a token plan, or that I'm an edge case burning through allowances at some unreasonable rate. I don't think that's it either. I run almost exclusively on 50%-off credits — the cheapest way to use the plan that's available to me. If the economics didn't work out even there, that's not a me problem, it's a plan problem.
What the mistake actually was
So here's the actual shape of the mistake: I bought a cheaper lane specifically to protect a more expensive one, and the cheaper lane couldn't absorb enough of my actual work to do that job. The purchase logic was simple going in — route the routine, high-volume stuff to Qwen, keep Claude for the parts that need it, and watch the Claude bill go down. That logic only holds if the Qwen allowance can actually carry the volume I'd be routing to it. It didn't hold.
Four days, mid-task
I felt this directly, not just as a line on an invoice. Mid-task, on August 25th, the plan told me: "Your token-plan 1-week quota has been exhausted." That's the message, verbatim, and it's worth sitting with because of when it landed — not at the end of a week of light use, but in the middle of something I was actively doing. The plan didn't come back until four days later. Four days is a long time for a lane you were counting on to just not be there, especially when the entire reason you set it up was to take pressure off somewhere else.
Capability and economics are different problems
That's the tell that this is an economics problem and not a capability problem. A model doesn't stop being good because its host plan ran dry. Qwen 3.8 was exactly as capable on August 29th as it was on August 24th. What changed was that I couldn't get to it, because I'd exceeded an allowance I didn't have good visibility into ahead of time. The model held up its end. The plan didn't hold up mine.
This matters because it's a really easy distinction to lose, and losing it leads to the wrong fix. If you conflate "the model disappointed me" with "the plan disappointed me," you either give up on a genuinely capable model for the wrong reason, or you keep buying the same kind of plan and expect a different result. Neither helps. The fix isn't a better model — I already have one. The fix is knowing, before you buy, whether the allowance on offer can actually carry your real usage, not your estimated or hoped-for usage.
The question I should have asked first
That's the part I got backward. I evaluated Qwen 3.8 on quality and decided it was good enough to trust with real work — which was the right call, it is good enough. What I didn't evaluate carefully enough was whether the plan's allowance matched how much of that real work I'd actually be sending its way, especially once I started leaning on it hard because it kept clearing the quality bar. The better the model looks, the more you route to it, and the faster you find the edge of an allowance you hadn't stress-tested.
The broader lesson, if there is one, is that "cheaper lane to protect an expensive lane" is only a strategy if you've checked the ceiling on the cheap lane against your real volume, not against a plan's marketing page or your own optimistic guess. A capable model you run out of partway through a task isn't a substitute for a lane you can actually keep using. It's a capable model you run out of. Those are different things, and the second one is the one that cost me here — not the model, the math.
What I would still tell someone
I'm not walking this back into "don't use Qwen." I'd still point someone toward Qwen 3.8 for real agentic work; my opinion of the model itself hasn't moved. What I'd tell them to do differently is separate the two questions before they buy anything: is the model good, and does the allowance match what I'll actually throw at it. I only asked the first question. The second one is the one that decides whether the purchase pays for itself, and it's the one I'd get right next time.