← Back to articles
November 9, 2026

Unreleased Qwen 4 Beats Claude Opus 5.5 at Game Development, at 15% of the Cost

Unreleased Qwen 4 checkpoints topped GameFeel Bench v1.4: 49.9 vs Opus 5.5's 47.2, at $0.40/task vs $2.60. One dev's benchmark; identity unconfirmed.

Unreleased Qwen 4 Beats Claude Opus 5.5 at Game Development, at 15% of the Cost

Alibaba has not released Qwen 4. But two early checkpoints labeled Qwen-4-Flash-Early and Qwen-4-27B-Early just appeared at the top of a game-development benchmark, ahead of Anthropic's Claude Opus 5.5, at a fraction of the price (GameFeel Bench v1.4 data, updated Nov 8, 2026).

The scores are real, published, and downloadable as CSV. What they mean is the interesting part: this is one developer's creativity benchmark, the models are unreleased, and Alibaba has not officially confirmed that the checkpoints are Qwen 4 at all. (⚠️ Partially verified: scores and prices verified against the benchmark's open CSV; model identity rests on the benchmark maintainer's labeling and tester reports, not an Alibaba announcement.)

The demo that started the week

On November 7, developer Loktar (@loktar00), a member of the House of Qwen ambassador program, posted a 30-second video of a 3D Colosseum packed with animated crowds and torchlight. His caption: "Qwen 4. ARE YOU NOT ENTERTAINED!!!"

The details he confirmed in the thread make it more interesting than the render itself (Loktar's replies, Nov 7):

  • One shot, zero tuning. "First shot! I ran it through a harness (Qwen Code) but since we only get so many tokens I stopped after the first output because it was so good."
  • Pure JavaScript and Three.js. No Blender, no MCP pipeline.
  • 69,000 tokens, a figure burned into the demo video's own caption ("Qwen 4 (early access) · Qwen Code · 69k tokens"). Developer Herlon Aguiar (@mxfp0) flagged it as the most impressive part: 3D scenes normally burn far more.
  • Thinking set to High, not the higher XHigh or Max tiers.

Loktar says the API came with no model name or specs. Asked whether it was Qwen4-27B, he answered: "I wish I knew, would be amazing if it was!" He's still testing: on November 9 he posted about his "limited Qwen 4 tokens" (@loktar00). (✅ Verified: all quotes read directly from the public X thread, Nov 9, 2026.)

What GameFeel Bench actually measures

GameFeel Bench is built and maintained by one developer, follistatin (@follistatindev, GitHub: kodasdream/gamefeel-benchmark). It is not a general-capability leaderboard like AA or Arena, and it does not pretend to be.

The methodology, from the project README:

  • 12 real but unpublished game projects (roguelikes, MMORPGs, single-player RPGs, survival/sandbox), chosen so no model can have seen them in training.
  • 7 subcategories: coding create, coding bugfix, gamefeel, creativity ideas, visuals, audio, and reverse-evaluation. Roughly 60,000 tasks in total.
  • Models are told to add 30 to 100 new items, mechanics and balance changes per game, then graded on whether the additions actually work and stay balanced.
  • Coding and reverse-evaluation are answer-key graded, with Opus 5.5 as the checker. Gamefeel and creativity are human-rated, 1 to 10 per task.

That last point cuts both ways: human rating catches things pass/fail tests miss, but it is one evaluator's taste, and the answer-key checker is itself a contestant. (✅ Verified: methodology quoted from the project README, Nov 9, 2026.)

The scores: Qwen 4 Flash first, Opus 5.5 second

The v1.4 overall table (average of all seven subcategories), verified row by row against the published CSV:

ModelEffortOverall scorePrice per task
Qwen-4-Flash-EarlyMax49.9$0.40
Claude Opus 5.5XHigh47.2$2.60
Claude Sonnet 5.5Max45.0$7.80
Qwen-4-27B-EarlyXHigh44.6$0.16

The maintainer announced the addition on November 8: "Qwen-4-Flash-Early (codename hawk) added to GameFeel Bench!" The post drew 413K views in a day. He also corrected an audio-score calculation error for the 27B entry in the same update, and says he paid per token for his own early-access API key (reply screenshot). (✅ Verified: announcement post and CSV agree; the CSV also lists the Flash checkpoint at 1M context, 240 tokens/sec, 23 minutes per task.)

The cost gap is the number developers will care about: $0.40 vs $2.60 per task is 6.5 times cheaper for the top score. The 27B checkpoint, listed as open-weight at 35GB, scored within 0.4 points of Sonnet 5.5 at $0.16 per task, one forty-eighth of Sonnet's $7.80. (✅ Verified: computed from the CSV's price_per_task_usd column.)

What Alibaba has actually confirmed

Very little, so far. The official Qwen account posted a teaser on November 6: "The becoming begins," over a poster quoting the classical Zhuangzi line about a giant fish in the northern darkness whose name is left blank. The missing word is "kun," a pun on Qwen. The post passed 2.6M views (@Alibaba_Qwen). (✅ Verified: post read directly, Nov 9, 2026.)

Chinese media report the family will include Max, Plus, Flash and 27B variants, and that Qwen Code repository updates in late October already referenced qwen4 and qwen4-max identifiers with million-token context support (QbitAI, Nov 9; Kuai Keji, Nov 8). (⚠️ Partially verified: consistent across outlets but tracing to the same leaks; no Alibaba announcement.)

The technical direction is public. In August, Alibaba open-sourced Qwen3.8-Flash-Next as an explicit preview of the Qwen 4 architecture: 125B total parameters with 6B active, hybrid sparse attention, gated residuals, N-gram embeddings and the Muon optimizer, trained at about 1/9 the cost of Qwen3.7-Plus (official blog; see our August analysis). If the early checkpoints hold up, that architecture is what produced them.

The caveats that belong next to the scores

Three limits are easy to lose when a chart goes viral:

  1. Identity is unconfirmed. The "Qwen-4" names come from the benchmark's labels and tester speculation. Loktar himself does not know which model he is running.
  2. One benchmark, one domain. GameFeel measures game-development creativity and balance. It says nothing about reasoning, math, or general coding, and its human-rated categories reflect one evaluator.
  3. Early checkpoints are not releases. Scores can move in either direction before launch, and the maintainer has already corrected one scoring error mid-week.

What survives the caveats is still notable: a Chinese lab's unreleased model is trading blows with Anthropic's best on a public, reproducible test, at 15% of the cost, and the data is open for anyone to check.

Sources

Verification summary: 5 claims ✅ verified against primary public sources (X threads, benchmark CSV/README, official teaser); 2 claims ⚠️ partially verified (model identity, family lineup — no Alibaba confirmation); 0 ❌ unverified; 0 🔴 contradictions. Overall: the benchmark data is solid and open; the "Qwen 4" framing remains an unconfirmed label until Alibaba ships.

Get notified

One email a week when new China AI news is published. Unsubscribe anytime.

Prefer a reader? Subscribe via RSS