Qwen 4 released the benchmarks and the price sheet together. The weights wait until December 20.

On November 11, Alibaba put the Qwen 4 and Qwen 4 Flash APIs online.
Word reached China that afternoon. A hands-on video on Bilibili titled "Qwen 4 is out! Open models face off against Claude Fable 5.2" had passed seven thousand views by the 13th. Those two words, "open models", are the most delicate part of this launch.
What they actually shipped
The price is $1 / $4 per million input and output tokens. None of the top five models in this year's flagship tier sits there. By Artificial Analysis figures, a full index task at the high reasoning setting costs about $2.94, while Fable 5.2 at the same setting costs roughly nine times that.
The score is 58, third on Intelligence Index v5. Ahead sit Gemini 4 Carbon (63), which Google has not fully opened up, and Fable 5.2 (65). Right behind sits the open-weights record from Kimi K3.1 Pro. On the sub-scores, Frontier Code 2.0 comes in at 47.9% and Frontier SWE v2 at 66.8%, which puts it between Opus 5.5 and Fable 5.2 on both agentic coding boards. On AA-Briefcase, the knowledge work board, it scores 1792 Elo, slightly above Fable 5.2 at 1785 and still short of Sonnet 5.5 at 1813.
On API design, Alibaba went the other way from Anthropic. temperature, top_p and top_k are all exposed, forced tool choice works, hybrid thinking switches per request, and you can set a thinking budget directly. Fable 5.2 keeps adaptive thinking always on and accepts none of those fields. For product teams this matters far more than two or three benchmark points, because thinking you can turn off means a bill you can predict.
Multimodality runs through one endpoint: text, images, audio and video all go in, with no separate Omni SKU. The launch demo was live interpretation across 19 languages with speech latency under 400 milliseconds. If that number survives replication, workstation-grade interpretation no longer needs a separate contract.
The benchmark spreading in Chinese circles needs an asterisk
The board that spread widest inside China is another one: GameFeel Bench.
According to a leaked test report, Qwen4-Flash scores 49.9, above Opus 5.5 xHigh at 47.2. Qwen4-27B scores 44.6, nearly level with Sonnet 5.5 at 45. Cost per task is $0.4 and $0.16, against $2.6 for Opus 5.5. If the 27B tier really reaches that level on a single 4090 or one Mac Studio, small and mid-sized teams will have to redo their self-hosting math.
The problem is the board itself. GameFeel Bench is an evaluation the Qwen team built for its own models. The task reads: given a game concept in natural language, produce a complete system with physics collision, state machines, economy values and interaction feedback. Questions are generated dynamically and cross-checked by humans, which the team says defeats memorization. I like the design idea. Running a whole system end to end is closer to real engineering than answering multiple-choice items. But the test set is not public and no third party has reproduced it, so the number is a vendor's own account for now. Wait until someone throws their own project at it before deciding how much to believe.
The weights are a date on a calendar, not a delivery
What actually distracts developers is this: the November 11 launch covers the API tier only. No download, no Hugging Face repo, no GGUF.
Alibaba promises to release Qwen 4 (1.2T total parameters, 56B active) and Qwen 4 Flash (30B total, 3.3B active) under Apache 2.0 by December 20. The date is specific, but it stays a box on a schedule. There is no file on disk yet.
The newest Qwen with public weights is still Qwen3.8-Flash-Next. Which means the best value model on that price sheet is not something you can run yourself today.
There is a positive precedent. Moonshot hit its last promised weight date without a single day of slippage. There is a negative one too. A release the community projected from a cadence chart for late October actually landed eleven days late. So put this promise in your calendar. Do not put it on a purchase order.
Whether to buy depends on whether you want to run it yourself
Put the three together and it becomes clear.
If you want cheap, stable and usable today, the Qwen 4 API is one of the best value picks this year, especially for batch jobs where thinking is off and spend stays inside a budget.
If you need compliance, offline operation and the model in your own data center, you cannot sign yet. You wait for December 20, and what arrives is a 1.2T set of weights. That size is not download-and-run. Finding the VRAM and assembling the inference stack will keep a team busy for two weeks.
As for that 49.9, I would treat it as a signal awaiting verification. Chinese models have genuinely caught up in the most expensive tier, and they hold the price low. That part is fact. "Beating Opus 5.5" is a claim that needs a third party to run it on their own questions first.
One cost item gets overlooked: migration. Folding multimodality into a single endpoint means work that used to run as three pipelines, image understanding through a vision model, speech through ASR plus TTS, video through frame extraction fed back into images, now happens in one call. One less pipeline means one less auth layer, one less rate limit config, one less person watching error logs. That math never reaches a financial report, but it decides whether three people can maintain four products at once.

What to watch next
December 20 is the first gate. If the weights do land that day, watch three things instead of the download count: how fast quantized versions follow, what context length fits on a single card, and how long framework support takes. Putting a 1.2T model into your own data center has never been a one-command job.
The second gate is third-party reproduction. If other labs are willing to test their own models on GameFeel Bench's "end-to-end runnable system" tasks, the board gains credibility. If a year from now only one company reports scores on it, it was launch material.
The third gate is the most practical: whether the price holds. $1/$4 is a customer acquisition price, and the record of the past two years suggests it is unlikely to stay put once share stabilizes. Hard-code that number into your budget model today and you may be rewriting it next year.
One more thing. This round of Chinese launches shares a pattern: put the API and the price sheet out first to win customers, hold the weights back for a second wave. For users, that means "open source" increasingly describes an installment plan rather than a delivery. That is not deception. It does call for changing how we read launch announcements.
Related articles
The Jensen Discount: What Nvidia Loses When Its Founder Walks Out
Nvidia lost roughly $600 billion in market value when its founder stepped down. The earnings did not change. The multiple did, and that says where the value really lived.
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.