Qwen 4 Leaks Early: Tops a Game-Dev Benchmark and Undercuts Opus 5.5 on Cost

Qwen 4 has not launched yet, but early test results are already circulating. A developer in the House of Qwen ambassador program was handed an API with no model name and no spec sheet, asked to try it out. What came back was a 3D rendering scene built in plain JavaScript and Three.js in a single generation — and the token bill was unusually small.

A Single Generation, Under 70,000 Tokens

The tester, Loktar, said the whole scene came from one shot: roughly 69,000 tokens, thinking mode set only to High rather than the higher XHigh or Max settings, and no prompt tuning at all. Rendering a 3D scene is normally one of the most token-hungry things you can ask a model to do, which makes the result notable on cost grounds alone.

He could not confirm the model was Qwen 4, but said it would be excellent news if it was.

The Benchmark: GameFeel Bench v1.4

A third-party leaderboard adds weight to the speculation. GameFeel Bench scores models on game development specifically — not just whether the code compiles, but whether the game feels right, looks right, and shows real design choices.

On the current v1.4 board, two suspected early Qwen 4 builds sit in the top four:

The Flash variant edges Opus 5.5 into second place, and the 27B build lands fourth within 0.4 points of Sonnet 5.5.

Cost: About 15% of Opus 5.5

The leaderboard's testers also disclosed per-task cost. Qwen-4-Flash-Early completed a run for roughly $0.40. Opus 5.5 came in at about $2.60. That puts the Flash build at roughly 15% of the cost — a gap of about 6.5x.

If a shipping version holds that ratio, it changes the arithmetic for developers who budget in tokens every day. The leaderboard's maintainer appears to have received test access on similar terms, with an API key provided but tokens paid out of pocket.

What We Know About the Lineup

Qwen has teased the release with a poster reading "The becoming begins" alongside a line from Zhuangzi about a fish in the northern sea — the character for that fish, kun, is a pun on Qwen. The lineup is expected to span Max, Plus, Flash, and 27B variants, covering both flagship capability and low-cost inference.

The architecture direction has also been previewed through Qwen 3.8 Flash Next, which introduced hybrid sparse attention, gated residual connections, N-gram embeddings, and the Muon optimizer. The stated goal is lower training and inference cost without giving up capability: 125B total parameters with only 6B activated per inference, and training cost around one-ninth of Qwen 3.7 Plus.

Two further signals point the same way. Late October code updates to Qwen Code included identifiers such as qwen4 and qwen4-max, along with million-token context adaptation. And the mystery test APIs have been reaching outside developers this month.

A Crowded Release Window

Qwen 4 is not the only model expected imminently. GPT-6.2 Astra has already been spotted in Codex usage statistics, Anthropic's Fable 5.5 is reportedly in staged rollout with Mythos 5.5 appearing in public page code, and Google's rumored Gemini 4 Carbon is said to open limited access this week. After months of "coming soon," the audience has grown skeptical — the next question is whether any of these vendors actually ships.