Alibaba Opens Free Qwen 4 Max Preview, Beats Claude Opus 5.5 in GameFeel Bench

Alibaba has rolled out a preview of Qwen 4 Max and opened free access to it through the chat.qwen.ai web chat. The model is positioned as a tool for programming and working with agents, exactly the tasks where the main race between labs is happening right now. Telegram channels are already calling it a "free monster," but let's separate fact from enthusiasm.

What is actually known

There are few facts. The model is called Qwen 4 Max Preview, access is free, and it is tuned for code and agentic scenarios. The only number from the first tests is its result in GameFeel Bench, a benchmark that evaluates game creation: Qwen beat Claude Opus 5.5 there. The source gives no details: no context window size, no API pricing, no information on open weights, and no results on the usual engineering tests. We won't make them up. The word Preview matters too: this is an early version, and the author of the original post wryly notes that you should use it while it's on offer.

Why the comparison with Opus 5.5 is shaky

The main caveat: Qwen was run at maximum effort, while Claude was run at medium. Effort, or reasoning budget, is how many "thinking" tokens a model can spend before answering. Imagine two programmers with the same task: one is given half an hour, the other five minutes. If the first does better, that doesn't mean he is stronger. Maximum effort usually improves results on hard tasks noticeably, but makes the answer more expensive and slower. A fair comparison is possible either at equal effort, or at equal cost or time per task.

There is a second problem: this is a single narrow benchmark. Games are not the same as a production backend, legacy refactoring, or long chains of edits in CI. Agentic tasks are especially revealing. An agent takes dozens of steps: it reads files, calls tools, edits code, runs tests. An error at an early step drags all the others down with it, so what matters is not polished demos but stability over a long run. A game benchmark measures that only partially.

What to do about it in practice

A free preview is a reason for a cheap experiment, not for changing your stack. A sensible plan looks like this:

1. Take 5–10 real tasks from your backlog: a production bug, a small feature, writing tests, a migration.
2. Run them through Qwen 4 Max Preview and your current model with an identical prompt.
3. Record not only quality but also response time, the number of revision iterations, and token usage, if possible.
4. If you compare against Claude, run it at both medium and high effort to see a fair picture.
5. Separately test a long task: does the model lose context, does it get stuck in loops?

And don't send proprietary code or keys to a free web chat. The source doesn't describe the data usage terms, so read the service's policy yourself.

For product managers, the takeaway is this: yet another strong free competitor puts pressure on market prices. Even if Qwen doesn't overtake Opus, it could become a cheap backup model for tasks of medium complexity. For founders: don't build a product on a preview. Limits, terms, and availability can change without warning.

Bottom line

This news is more of a "worth trying" story than an "everything has changed" one. One number under asymmetric settings is a weak argument, but free access makes checking nearly costless. Wait for independent tests across several benchmarks, look at API pricing and terms, and only then decide whether to let Qwen into your workflow.