Independent · Source-cited · No sponsored rankingsSources linked in every article
AIToolRanked
AI News · 5 min read

Qwen4-Max is $1.50 in and $4.50 out

On November 16, 2026 Alibaba opened qwen4-max at $1.50 and $4.50 per million tokens. Weights stay closed. The Work360 savings are Alibaba's own test.

RA
· Founder, AIToolRanked
A dark office at night with a long glass table, one lit desk lamp, and a closed notebook facing a city skyline
On this page

What Alibaba opened on November 16

On November 16, 2026, Alibaba Cloud's model release notes ↗ added qwen4-max for the International deployment scope. Qwen's launch post, Qwen4-Max: Your New Teammate Has Arrived ↗, is dated 2026-11-16 10:00 UTC+8 in the post record. That post says the model is on QwenCloud through the Token Plan or the API, and also in QwenWork, Qoder, and QwenStudio.

The picture above is ours. We have not called the model.

Qwen says this is the first Qwen4 model: a sparse mixture-of-experts with a 2.4-trillion-parameter backbone, 71 billion parameters active per token, plus 328 billion N-gram embedding parameters, native vision, and a 1 million token context. Those counts are Qwen's. There is no weights file.

This is a different ID from the November 11 preview. The preview page ↗, updated November 16, still says qwen4-max-preview is Token Plan only and still retires it at 24:00 on November 25, 2026 (UTC+8). Our preview note is the subscription-key piece. The pay-as-you-go ID is qwen4-max.

The blog's last paragraph says smaller models "are coming to open source, built on the same lineage and trained with the same generation of techniques." No sizes, no date. Max is not a download.

What the price page charges, and which cap to use

Model Studio pricing ↗, last updated November 16, 2026, has no input-length tiers for qwen4-max. Figures are per million tokens.

ScopeInputOutputCache read
Singapore, International$1.50$4.50$0.075
Beijing, and Global$1.413$4.24$0.071

Global on that page is Hong Kong, Frankfurt, US (Virginia), and Tokyo. Singapore cache writes are $1.875 (five minutes) and $3 (one hour). Singapore lists a 1 million token free quota for 90 days. Other regions have none. On the same Singapore table, qwen3.8-max is $2 and $6.

Qwen's spec page ↗ is the cap to use. Context 1,000,000. Non-thinking output 131,072. Thinking and the answer share 262,144. Thinking defaults to reasoning_effort high; the preview still defaults to xhigh. Input is text, image, and video. Tools include function calling, structured output, and web search, on OpenAI, Anthropic, and DashScope APIs. Batch, fine-tuning, and deployment are unsupported. Beijing's price table still prints a Batch File rate of $0.707 / $2.12. That line does not turn batch on.

The marketplace card ↗ still prints max output 393K, the preview's 393,216. Do not size qwen4-max to that. Vercel ↗ (November 15) lists alibaba/qwen4-max at provider pricing with no markup.

What Qwen claims, and what an outside index measured

Every Qwen score below is self-reported. We did not run the tests.

On Qwen's internal Work360 eval, the blog says Qwen4-Max scored 52.69 against 47.74 for Qwen3.8-Max. Token use fell about 24% at xhigh and 35% at high. On Terminal-Bench 4.0, Qwen says token use fell about 48% and 49%. Average cost per Work360 task fell from $4.43 to $1.67, which Qwen calls 62%. That workload is theirs.

The blog says cybersecurity data was filtered by task, defensive work kept and offensive attack data removed, with a CBRN safety pass in post-training. SCMP ↗ reports that filtering and says Qwen 5 will be scaled to 5–10 trillion parameters. The Straits Times ↗ reports that Alibaba said Qwen4-Max beat Claude Fable 5.1 and Kimi K3.1 Pro on coding benchmarks. It does not name the tests.

Artificial Analysis ↗, November 16, scores Qwen4-Max at 43 on its Intelligence Index at xhigh: 9 points ahead of Qwen3.8 Max (34) and 3 behind Kimi K3.1 Pro at max (46). AA's cost per index task is $3.27, against $10.02 and $3.68, about 11% under Kimi. On Terminal-Bench 4.0, AA has Qwen4-Max at 48% and Kimi at 67%. That is not a coding win over Kimi. AA also puts Qwen3.8 Max cache hits at $0.25. Alibaba's page prints $0.075 for Qwen4-Max. The $0.25 is AA's.

What people are saying

  • "it's $1.50 / $4.50 per million tokens, about 4x cheaper than Opus 5.5. Hopefully this is a lot better than the free Qwen4 Max preview I tested about a week ago, because that was pretty awful." — Layton Gott ↗ (@Layton_Gott), November 16. Our November 15 Gemini note recorded Opus 5.5 at $4 and $20 on Anthropic's page that day. On those rates, $20 / $4.50 is about 4.4 on output. Input is $1.50 against $4.

  • "Pretty sure Max is more a teacher model for the new smaller models they distill from it" — Danny Hallwood ↗ (@stepbystepnomad), November 16

  • "On Terminal-Bench 4.0: Kimi K3.1 Pro: 67% Qwen4-Max: 48%" — FHILY ↗ (@Oluwaphilemon1), November 16. That is the Artificial Analysis row.

  • "the most important sentence in today's technical blog is at the very end!" — David Hendrickson ↗ (@TeksEdge), November 16, on the open-source line

Reddit was not reachable for this story.

Our take

Use it when the token bill is the constraint and a listed region is acceptable. Singapore at $1.50 and $4.50 sits under Qwen3.8-Max at $2 and $6, and under the Opus 5.5 rates in that November 15 note. Gemini 4 Hydrogen is $1.50 and $7.50 there, so output costs more per token. Qwen's 62% is self-reported. AA's index still puts Qwen behind Kimi.

Do not treat Max as the local model. The blog promises smaller open models and names no date. Until a file exists, a machine you control starts with Kimi K3.1 Pro or LongCat-2.5. DeepSeek V4.1 Flash Ultrafast is a speed tier, a different product.

Size the job to 262,144 with thinking on, or 131,072 with it off. The 393K figure is the preview, and that ID still ends November 25. Leaving reasoning_effort unset is high, not last week's xhigh default.

Who should try it: teams on Claude or OpenAI list rates for long agent and document work, calling Singapore, Beijing, or the Global cities above, who will measure their own task. Who should skip it: anyone who needs weights, batch, or a Terminal-Bench 4.0 lead over Kimi. The map is the open-source LLM comparison and the best AI for coding guide.

More from the news desk · About Rai Ansar