Alibaba’s Qwen 4 benchmark leak put Qwen4-Flash at 49.9 points on GameFeel Bench, ahead of Opus5.5xHigh at 47.2. The result matters because it places Qwen 4 ahead on a game development creativity test that compares model output on a specialized task, not on broad chat alone.
Qwen4-Flash and GameFeel Bench
GameFeel Bench is the benchmark at the center of this result, and the score gap is straightforward: 49.9 against 47.2. That means Qwen4-Flash edged Opus5.5xHigh on the reported test, while Qwen4-27B also posted 44.6 points and Sonnet5.5 reached 45 points.
The leaked numbers suggest that Alibaba is not just chasing raw quality. The reported execution cost for Qwen4-Flash was about $0.4 for an equivalent task, while Opus5.5 was about $2.6, so the cheaper option in the leak is also the higher scorer on this benchmark.
Alibaba and the cost gap
The math behind that cost spread is blunt. $0.4 versus $2.6 is a difference of $2.2 per equivalent task, and Qwen4-Flash therefore comes in at roughly one-sixth the reported cost of Opus5.5. Qwen4-27B was even cheaper at $0.16, which gives Alibaba a second price point to point to if the company wants to sell different tiers of performance.
That matters for developers who care about output quality and inference bills at the same time. A model that can stay competitive on a creativity benchmark while costing less to run is easier to justify in production, especially when a project needs repeated calls rather than one-off demos.
Yunchi Conference in Hangzhou
The leak did not appear in a vacuum. On September 22, Alibaba said at the Yunchi Conference in Hangzhou that Qwen4 had entered training, and Liu Dayi Heng said, “scaling is an important path toward ASI.”
That earlier message now sits beside a harder question. Qwen 4 is being presented as a breakthrough, but the benchmark results are described as leaked information rather than an official release, and Alibaba has not laid out the full public launch package in the facts provided here.
