Morning Edition · Monday, November 16, 2026Published at 1:19 AM EST · New York
On Alibaba's in-house Work360 benchmark the model gains about five points over Qwen3.8-Max, but the only independent index so far scores it at 43, against 59 for Anthropic's Claude Opus 5.5.

Alibaba's Qwen team released Qwen4-Max, a mixture-of-experts model with native vision and a one-million-token context window. According to the launch post, the model has a 2.4-trillion-parameter backbone with about 71 billion parameters active per token, plus 328 billion parameters in a separate N-gram embedding table. Artificial Analysis lists the release date as November 16.
The architecture changes compute per token more than it changes model size. Qwen3.8-Max, which Alibaba opened to the public on August 3, also had 2.4 trillion total parameters, with 95 billion active. Qwen4-Max therefore keeps the total size and cuts active parameters by about a quarter. It also adds embedding parameters, which the model reads by lookup rather than through per-token matrix multiplication. Alibaba's announcement on X says the model beats Qwen3.8-Max while using up to about 50 percent fewer tokens, and that average cost per task falls by more than 60 percent.
The headline numbers come from Alibaba. On Work360, its in-house benchmark, Qwen4-Max scores 52.69 against 47.74 for Qwen3.8-Max, and Alibaba says the average task cost falls to $1.67. A developer analysis points out that no independent party has verified these benchmarks. The only third-party figure available is the Artificial Analysis Intelligence Index, which scores the model at 43. The median for comparable models is 41. A comparison by Kingy AI puts Claude Opus 5.5 at 59 on the same index. Artificial Analysis also says Alibaba has not disclosed the model size, even though Alibaba's own blog gives the 2.4-trillion figure.
The international list price is $1.50 per million input tokens and $4.50 per million output tokens, with cached reads at $0.075. Qwen3.8-Max cost $2 and $6, so both rates fall by 25 percent. The cut still leaves Qwen4-Max above Meta's Muse Spark 1.1, which has been in public preview since July at $1.25 input and $4.25 output. FourWeekMBA calculates that the output price is 91 percent below OpenAI's GPT-6 Astra.
The weights are closed. Artificial Analysis says they are not publicly available. Alibaba says open weights for smaller Qwen4 models will follow but has given no date. Until last month, Chinese-language trackers reported that Alibaba had previewed the Qwen4 family at its Apsara conference in September without publishing any specifications.
Part of a tracked trend
Frontier Model Price War
Frontier API vendors increasingly compete on strategic price-cutting rather than pure capability, repeatedly launching or repricing models below prevailing rates to grab share and compressing industry inference margins; expect recurring below-rival pricing moves.
Start a discussion in Townsquare.
More from this edition
If this framing holds, Alibaba Cloud gains. A lower cost per agent task draws enterprise inference spending to its closed application programming interface (API), and Western vendors of mid-tier agent APIs take the pricing pressure.
Chinese-language outlets confirm the November 16 launch and the Work360 scores of 52.69 and 47.74. However, Sina and SMZDM put the per-task cost cut at 58 percent (26.58 yuan to 11.16 yuan), not more than 60 percent, and that figure comes only from Alibaba's own harness. The edition also leaves out a second third-party data point, a reported 14th place on the Arena agent leaderboard that Chinese coverage calls first among domestic models, and the desk could not independently retrieve the Artificial Analysis score of 43.
An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.
What this means
Alibaba is competing on cost per completed agent task, not on rank at the top of the leaderboard. It cut active parameters and claims large token savings, which lowers its own serving cost, and it passed part of that saving into a 25 percent price cut. Two groups are exposed: Western vendors selling mid-tier agent APIs, and Alibaba Cloud's inference margins. A 16-point gap to Opus 5.5 on an independent index means the model does not threaten closed frontier pricing at the top tier. Many observers assume that "Chinese model" means "open weights", and this release cuts against that assumption: Alibaba kept its flagship behind an API, as Western labs do, and promised open weights only for smaller models. The likelier outcome is that pricing pressure lands on the mid-tier within weeks, through cheaper "flash" or "mini" tiers from US vendors, not through flagship cuts. That outcome depends on independent agent benchmarks confirming that the token savings survive harnesses Alibaba does not control. If third-party runs show cost per task close to Qwen3.8-Max, the price cut is a list-price change and not an efficiency gain.
What to watch
Observations to monitor, not financial advice.
Synthesized from: Qwen (via Hacker News) · Meta AI
Comments
0No comments yet.