❯ Early Qwen4 models reach a game development benchmark, with Flash at 49.9 and 27B’s audio score corrected

Two early models enter an independent testOn November 8, GameFeel Bench added Qwen-4-Flash-Early. The downloadable leaderboard shows 49.9 for Flash at Max effort and 47.2 for Claude Opus 5.5 at XHigh. An audio scoring correction moved Qwen-4-27B-Early from 47 to 44.6 at XHigh, close to Claude Sonnet 5.5’s 45 at Max.

The task is to improve the feel of an existing gameThe benchmark uses 12 unreleased games and asks models to add content, fix issues, or rebalance gameplay. It considers ideas, game feel, visuals, and audio alongside code. The author says that humans rate some categories and know which model produced each result. Flash’s showing in gameplay and ideas helped lift its overall score, while the results for any one task depend on the project and the scoring criteria.

The cost figures come from the evaluator’s own billThe author said he paid about $0.40 per task for Flash and $0.16 for 27B on an early-access API. The benchmark’s $2.60 for Opus 5.5 uses subscription billing. Those numbers reflect the test settings and accounts used. Public API rates, request volumes in a different project, and the ability to run a model locally will depend on what is released.

Public release details are still openAlibaba Cloud said in September that Qwen4 was in training. The benchmark shows that an evaluator could use early versions, while Alibaba has yet to publicly confirm the full lineup, a release date, or production pricing. The tests give game teams a look at what a smaller model can do on specific creative tasks. Repeating the test on the public versions will require access to those versions.

▮ SIGNALA lower per-task cost can give a game team room to try more mechanics in parallel, provided the model can produce something playable in the team’s own codebase.