Qwen4 has not been officially released yet, but the benchmark scores are out first.
The leaked results come from GameFeel Bench, a benchmark that specifically evaluates the creativity of game development to see how good the small games made by the model are:
| Model | Score | cost per task |
|---|---|---|
| Qwen4-Flash | 49.9 | About $0.40 |
| Opus 5.5 xHigh | 47.2 | About $2.60 |
| Sonnet 5.5 | 45 | not disclosed |
| Qwen4-27B | 44.6 | About $0.16 |
If true, this is the first domestic model to overtake the Opus 5.5 in this evaluation, and the overtake is the lightweight Flash version, which costs only about one-sixth of the other model. 27B, which can be deployed locally, is almost equal to Sonnet 5.5, making it more attractive to teams that want to deploy it themselves.
You also have to pour cold water on it. GameFeel Bench is a very narrow vertical evaluation. The subjective component of the game feel indicator is not small. Being ranked first in a single item cannot explain the overall ability. The leaked data cannot verify the test conditions. Opus uses the xHigh setting, and it is not clear what settings Qwen used.
The family lineup is basically clear: there are at least four models: Flash, 27B, Plus, and Max. The release node promised by the Qianwen team is just around the corner. When the time comes, it will be compared head-to-head with GPT-6.1 and Opus 5.5, and the real results will be considered.
If you want to get familiar with this series first, you can check out Tongyi Qwen3 and Qwen Image on the site.