Alibaba's early Qwen 4-era model beat Claude Opus 5.5 on a game-building benchmark — 34 wins to 10 — and the result is blowing up on X today. The test is GoldieBench: the same fixed prompt given to both models, one shot, one HTML file out, across 50 tasks — every game actually playtested. Qwen3.8-Max-Preview averaged 8.10 out of 10. Opus 5.5 averaged 7.60. The gap was widest where atmosphere mattered. Qwen took a synthwave 3D shooter 8.7 to 5.8, a twilight village 8.4 to 5.5, and a top-down action-RPG 8.2 to 5.5. It won 8 gold medals to Opus's zero. Opus still won its own battles — and one was a blowout. Its ocean scene scored 8.4 against Qwen's 2.0, the biggest single-task gap in the whole test. The catch: this is one independent bench, and the rules were not identical. Qwen's game tasks ran inside Qoder's agent mode with up to two QA fix rounds — a vision judge fed console errors back and Qwen edited its own file. Opus ran one-shot. And a separate test published November 7 went the other way: on a Sokoban build judged by GPT-6 Astra, Opus scored 77% to Qwen's 70% — in about 4 minutes, against Qwen's 31. The model itself is still preview-only: Alibaba's roughly 2.4-trillion-parameter flagship, available only inside Qoder, with no public API yet. Qwen 4 itself is still in training. But the direction is hard to miss. The open-weight challenger is now winning the creative-build tests — and the gap between a preview model and the closed frontier just got a lot smaller. BTC sits near $75,980, up about 0.1% on the day.